International Conference on Digital Transformation and Applications (ICDXA) Proceeding

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Figure 6. Results of Face-api.js, Tensorflow and MediaPipe Face Detection in Machine C 5.0 CONCLUSION In conclusion, through the analysis of Tensorflow Handpose and MediaPipe Hands, in terms of accuracy, Tensorflow Handpose is higher than MediaPipe Hands. The accuracy of Tensorflow Handpose is as high as 99.4%, while the accuracy of MediaPipe Hands is around 95.4%. However, in terms of FPS, MediaPipe Hands is more stable in terms of lower hardware specification requirements (i.e., machine A) or higher hardware specification requirements (i.e., machine C) than Tensorflow Handpose. Through the research on hand gesture detection, in order to have high efficiency on low-specification hardware, it is recommended to use MediaPipe Hands, on the contrary, Tensorflow Handpose is more suitable for high-specification hardware. In addition, the face-api.js provides the best performance in executing the face detection function compared to TensorFlow and MediaPipe, therefore the library is used for implementing the face detection in a virtual classroom system. Although there is a difference of average time execution for face detection among each of the libraries in different machines, each of the libraries in different machines shows similar trends where the face-api.js provides the highest performance compared to others. According to the research, face-api.js is recommended for high efficiency on low specification hardware, whereas MediaPipe is more suitable for high specification hardware. The performance of the libraries may be affected based on the programming languages, therefore a comparison for the performance of the libraries in different programming languages such as python versus javascript can be carried out in the future. Furthermore, in the future, in addition to detecting hand landmarks, another study can be proposed to enhance hand detection in virtual classroom systems by implementing hand gesture detection and recognition in virtual classroom systems. Besides, there are many face detection libraries which can be used for implementing the project. Hence, the better library can be used for 92 of 225 ICDXA/2021/08 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 analysis and implemented into the virtual classroom system in order to improve the performance of the system in the future. 6.0 ACKNOWLEDGMENTS With our sincerest greetings, we would like to express our gratitude to all those who provided great help, moral assistance and cooperation to our research team. We are grateful to the Faculty of Computing and Information Technology (FOCS) of Tunku Abdul Rahman University College (TARUC) for providing the facilities at the Center of Computational Intelligence (Center of Computational Intelligence, 2020) to make the project successfully implemented with preliminary results, as shown in this paper. REFERENCES AIT Staff Writer 2021, Role of Artificial Intelligence in Online Education, AiThority, viewed 20 September 2021, <https://aithority.com/technology/education-and-research/role-of- artificial-intelligence-in-online-education/>. Akshara J., Akshay J. T. L., Krishna Y. (2017), Automated attendance system using face recognition, in ‘International Research Journal of Engineering and Technology (IRJET)’, Vol. 4, pp. 1467–1471. Arun K., Sudesh V. K., Amar. P. Z., Nikhil. D. B. & Chetan. J. B. (2017), Attendance system using face recognition and class monitoring system, in ‘International Journal on Recent and Innovation Trends in Computing and Communication’, Vol. 5, pp. 273–276. Center of Computational Intelligence (2020), Tunku Abdul Rahman Universiti College, viewed 10 Sept 2021,<https://www.tarc.edu.my/focs/research/cci/>. Chintalapati, S., Raghunadh, M. V. (2013), Attendance system using face recognition and class monitoring system, in ‘International Journal on Recent and Innovation Trends in Computing and Communication’. Howard, C. (2018), ‘Face recognition based automated student system’. Bachelor of Engineering (Hons) Electronic Engineering. J. Valera, J. Valera and Y. Gelogo, 2015, \"A Review on Facial Recognition for Online Learning Authentication,\" 2015 8th International Conference on Bio-Science and Bio- Technology (BSBT), pp. 16-19, doi: 10.1109/BSBT.2015.15. Kanan, C. and Cottrell, G., 2012. Color-to-Grayscale: Does the Method Matter in Image Recognition?. PLoS ONE, vol. 7, no. 1, pp. 29740. Olszewska, J. I. (2021), The virtual classroom: A new cyber physical system, in ‘2021 IEEE 19th World Symposium on Applied Machine Intelligence and Informatics (SAMI)’, pp. 000187-000192. Pratiksha M. Patel, 2016, Contrast Enhancement of Images and videos using Histogram Equalization. International Journal on Recent and Innovation Trends in Computing and Communication, vol. 4, no. 11. Racheva, V 2018, What is a virtual classroom, VEDAMO, Vedamo, viewed 20 September 2021, <https://www.vedamo.com/knowledge/what-is-virtual-classroom/>. Rapanta, C, Botturi, L, Goodyear, P, Guàrdia, L & Koole, M 2020, ‘Online University Teaching During and After the Covid-19 Crisis: Refocusing Teacher Presence and Learning Activity’, Postdigital Science and Education, vol. 2, no. 3, pp. 923–945, viewed 20 September 2021, <https://link.springer.com/article/10.1007/s42438-020- 00155-y>. 93 of 225 ICDXA/2021/08 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Ren, Z., Meng, J., Yuan, J. & Zhang, Z. (2011), Robust hand gesture recognition with kinect sensor, pp. 759–760. Singh, S., Rastogi, R. & Sharma, P. S. (2015), Automatic lecture attendance system using face recognition, in ‘MATRIX Academic International Online Journal of Engineering and Technology’, Vol. III, pp. 36–40. Yufei, L, Saleh, S, Jiahui, H, Mohamad, S & Abdullah, S, 2020, ‘Review of the Application of Artificial Intelligence in Education’, International Journal of Innovation, Creativity and Change. www.ijicc.net, vol. 12, no. 8, p. 548 - 555, viewed 20 September 2021, <https://www.ijicc.net/images/vol12/iss8/12850_Yufei_2020_E_R.pdf>. Zhang, F., Bazarevsky, V., Vakunov, A., Thachenka, A., Sung, G., Chang, C.-L. & Grundmann, M. (2020), ‘Mediapipe hands: On-device real time hand tracking’. 94 of 225 ICDXA/2021/08 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 INTERACTIVE DASHBOARD WITH VISUAL SENSING AND ZERO-SHOT LEARNING CAPABILITIES Wen Lin Yong1, Jun-Kit Chaw2* and Yiqi Tew1 1 Faculty of Computing and Information Technology, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia 2 Institude of IR4.0, Universiti Kebangsaan Malaysia, 43600 Bangi, Malaysia *Corresponding author: [email protected] ABSTRACT These days, technology is growing rapidly, and the market has been introduced with lots of fascinating ways to interact with computers. The advancement of deep learning models and hardware technology also enables more applications with fancy features to be built. The importance of hand gesture recognition has increased due to the prevalence of touchless applications. However, developing an efficient recognition system needs to overcome the challenges of hand segmentation, local hand shape representation, global body configuration representation, and a gesture sequence model. This paper proposed an interactive dashboard that could react to hand gestures. This is also an initiative of the Tunku Abdul Rahman University College (TAR UC) Smart Campus project. Deep learning models were investigated in this research and the optimal model was selected for the dashboard. In addition, 20BN Jester Dataset was used for the dashboard development. To set up a more user-friendly dashboard, the data communication stream between the captured input stream and commands among the devices will also be studied. As to achieve higher responsiveness from the dashboard, evaluation on data communication protocols which were used to pass the input data included in the study. Keywords: Computer Vision, Human-Computer Interaction (HCI), Gesture Detection, Real-time systems, Feature Extraction 1.0 INTRODUCTION Human-Computer Interaction (HCI) is a study that proposes to discover the ways humans interact with computers. Due to the rapid growth in computer technology, this study has come up with various types of methods for the interaction between humans and computers. One of the well-known practices is the visual-based HCI which requires cameras as the input devices instead of computer mouse and keyboards. This kind of HCI tends to achieve a touchless user interface by capturing the user's facial expressions, human gestures, and more as the input data to control the computer. Although this field has been broadly studied in the computer vision field, the methods proposed in the market show vulnerability in the outdoor environment. The results produced by the gesture recognition algorithms which have been published are also unable to guarantee absolute precision (Chakraborty et al., 2017). The development of the interactive dashboard in this research aimed to achieve high performance in responding to user’s commands which were performed using gestures movements. Thus, the gesture recognition algorithms will be studied and evaluated to achieve 95 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 real-time capability in returning outcomes for the gesture movements detected. Aside from that, to obtain higher usability for the interactive dashboard, the video streaming protocols will be determined for faster transmission of gesture data captured from the input devices. Every individual in this world is known to act differently in their way. Even though a lot of defined classes and algorithms for human recognition systems have been introduced in the market, there might be chances of miscalculating poses while employing the practices for their use. To tune-fining the best gesture weights, there might require large gesture datasets. Besides, gesture recognition performance might be affected while implementing in a real- world environment due to external factors such as visual occlusions and more that affect the captured gesture data title. The user experience while making use of the dashboard might be affected by the sensitivity of the interactive dashboard. One of the factors that might affect the interactive dashboard irritability is the transmission of gesture videos data through the IP cameras and network video recorders. While there is a delay in the video transmission, the system might not achieve responsive controls towards the dashboard. Moreover, the model used to process the gesture will also affect the interactive dashboard reactivity (Gokul et al., 2019). 2.0 LITERATURE STUDIES We have studied two essential elements to be implemented in an interactive dashboard: human gestures and communication protocols to be utilized for HCI purposes. 2.1 Gesture recognition with deep learning models The interactive dashboard is presumed to be set out in the i2hub dashboard display which employed the 3x3 video wall solution and using gestures to control the dashboard seems to be the finest way to intercommunicate with the dashboard. Therefore, this project is anticipated to support gesture recognition functionalities. In this project, there will be making use of the RGB cameras as the input devices for capturing the gesture data. Although there are other methods enacted for better preciseness or faster pace in detecting the gestures, there will be some restrictions while implementing the methods in a real-world environment. Although utilizing wired gloves for input devices is known to be producing results explicitly and rapidly, there will be cutbacks on data collection as more gestures to be detected at one time requires more gloves which ends up costly for equipping more devices. Besides, it is inconvenient to make use of further distances due to wired restrictions. 3D hand key points location is not required on many devices; however, this method requires a longer time in reckoning the hand key points which deprived smooth interaction between the users and the dashboard (Köpüklü et al., 2019). We consider two deep learning models to be utilized for detecting gestures: Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) approaches. CNN is recognized to make use of object recognition with the ability to learn features in spatial data. By using the CNN model, gesture recognition can be performed in a fast behavior with high precision in the results (Guo et al., 2019). However, CNN models are incapable of learning temporal data which is crucial in continuous hand gesture recognition. Thence, 3DCNN is proposed to acquired features extraction on 3D data (eg. gesture movements videos data) which allows detecting dynamic gestures movements. Besides, recurrent neural networks (RNN) which are known for processing sequential data are taken into consideration to incorporate with the CNN model to process longer continuous gesture data (Chakraborty et al., 2017). Convolutional Gestures are required to be captured in a 96 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 continuous form to predict what actions have been done by the users (such as swipe up). To detect continuous gesture movements, recurrent neural networks (RNN) have been taken into consideration. However, Convolutional LSTM will be employed instead of the LSTM due to the LSTM model not acquiring temporal data while doing calculations (Shi, et al., 2015). 2.2 Data communication protocols Mentioned in the future expectations for HCI by Gaurav Sinha and fellow researchers, higher bandwidth interaction which denotes the acceleration of the interaction between human and computer should be premeditated to induce the sensibility of the interaction significantly (Sinha, Shahi, and Shankar, 2010). To produce high satisfactory HCI capability to the users, responsiveness from the dashboard will be concentrated in the development. In this project, IP cameras will be adopted as the input devices to capture the gesture instruction from the users, then the captured data will then be forwarded to the dashboard to do further processing and response. While there is communication between the devices such as the transmission of the input data from the camera to the dashboard, the communication protocols are required to enact and specified during the development of the dashboard. (Miroslav, 2018) In furtherance of the responsiveness of the dashboard, the speed of transferring the gesture data from the camera to the dashboard should be stimulated so that the dashboard can run the gesture recognition as soon as possible. Our scope of work covers the exploration of the video streaming protocols used to allow communication (e.g., passing media/files) between devices appointed and employment of the existing model/algorithms into the dashboard implemented to allow direct human interaction and controls to the dashboard. This work aims to develop an interactive dashboard with gesture-navigated effects based on artificial intelligence human posture detection which brings great convenience, flexibility, and efficiency by meeting the benchmarks listed below: 1. To attain high accuracy results in gesture recognition results. a. The gesture recognition algorithms and methods applied should be able to achieve high precision results to secure the usability of the interactive dashboard. By acquiring exact gesture recognition, the dashboard will be able to react with proper responses according to what users perform. 2. To achieve high responsiveness from the interactive dashboard. a. The video streaming protocols which have been employed should be able to transmit the gesture data at a higher speed to reduce the delay of video streaming to the system for gesture recognition processing. b. The suitable deep learning models used to perform the gesture recognition shall be implemented to accomplish the real-time capability of the interactive dashboard. 97 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 3.0 RESEARCH DESIGN AND METHODS Figure 1. Flowchart of overview interactive dashboard. Figure 1 illustrates the program flow for the interaction dashboard developed. The aim of detecting the human posture for direct interaction between humans and the dashboard, thus, some poses such as swipe left and right which are used to control the object movements inside the dashboard will be mainly focused on. As mentioned above, the posture might perform differently by every individual, hence the 20BN Jester dataset with a large amount of significant gesture data will be utilized in the development of the dashboard. Additionally, Zero-shot learning (ZSL) is expected to be implemented in this study to recognize undefined gesture classes by any chance. Besides, suitable deep learning models will be defined in this study to establish a well- functioned interactive dashboard. By applying the proposed models, human gesture recognition is also expected to be carried out in a real-time manner to provide a smooth HCI between users and the dashboard developed. Figure 2 shows the overall process of video retrieval in a proposed video streaming module. To capture human gesture movements, IP cameras and NVRs are employed to capture the postures that are performed by each individual. While human gesture recognition is proposed to perform in a real-time manner, the gesture data captured by those devices are expected to be transferred between those devices as quickly as possible to bring out a responsive and low latency interaction between the individuals and the dashboard. Then, there will be studies of video streaming protocols and deep learning models to deliver a better performance in the communication between those devices. 98 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Figure 2. Flowchart of video retrieval in video streaming module A convolutional HIKvision's IP cameras and Network Video Recorders (NVR) are used in this research. As HIKvision’s network cameras provide predefined video streaming protocols, settings on using which protocols can be done for specific needs on the video streaming. The protocols provided by the HIKvision products are TCP, UDP, HTTP, and MULTICAST. Every live streaming protocol achieves different goals: ● TCP: Compromising video stream quality and reducing packet loss problems, however, the real-time streaming will have delays. ● UDP: Support real-time video and audio streaming. ● HTTP: Similar to the TCP protocol, yet not required ports specifications ● MULTICAST: Establish multicast group addresses and provide stream acquisition by multiple users simultaneously. To obtain the video stream at a faster pace from the cameras to the dashboard, the UDP protocol will be selected at the moment (Hikvision, 2019). In terms of the experimental setup, we utilize the Jester dataset published by the Twenty Billions Neurons to train the gesture recognition model. This dataset contains an enormous amount of gesture movements video data captured from numerous actors. The gestures which acted in those videos are proposed to provide for public use in developing HCI (such as swipe up and more). Researchers who are interested in developing gesture recognition systems are also encouraged to make use of this dataset to construct the best solution for gesture recognition capability (Materzynska et al., 2019). 4.0 RESULTS AND DISCUSSIONS Before the construction of the interactive dashboard, the gesture recognition approach was tested and the results (gesture commands detected) were able to return. The approach was obtained through experimentation from other researchers’ studies. Figure 3 shows the test screens of the gesture recognition implementation in the flask application which is planned to be used for the dashboard set up. A swipe left gesture is performed and this gesture is expected to change to another video stream, as captured in Figure 3 (left). After a few seconds, the program returns the gesture results and prints them in the command prompt screen which shows in Figure 3 (right). Another few seconds passed, the gesture result successfully passed to the controlling section, and action was carried out. Then, the video stream has changed to another in Figure 3 (bottom). 99 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 (left) Swipe Left action is performed, (right) Delay response in results. (bottom) Command occurred and the stream has changed Figure 3. Sample output of testing on gesture recognition capabilities Besides, exploration on the retrieval of the previously recorded video stream for further development on the commands that needed to change the stream from real-time stream to the recorded stream such as rewinding and forwarding the video stream controls. Figure 4 shows that it was able to retrieve the past recorded video information from the Hikvision NVR. Figure 4. Past recordings information retrieved However, in Figure 3 the whole testing for the program is not running smoothly, the video is lagging, and the program takes time to return the result. This might be because all the 100 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 functions run in one individual function, so it requires the previous task to be done only to proceed to the next task. Therefore, tasks have been separated into several functions’ definitions. Figure 5 shows the output after the tasks have been separated into individual functions. The screen output in the right shows the results with the detected gesture, returns values 5 and “None” mean that there is no gesture detected, the values 4 and “Next” means that “Swipe Left” gesture detected, the values 3 and “previous” means that “Swipe Right” gesture detected. Yet while running the application, the dashboard keeps on showing a loading screen and is unable to prompt out the dashboard. Meanwhile, gesture recognition is running normally. This might be due to the functions having separated, they were not running concurrently. This means that it remains the same that it requires to wait for one task to be done only then proceed to the next task. Figure 5. Only run the long-run task (gesture recognition) and dashboard unable to be prompted To eliminate the blocking of calling other functions that are caused by the long-run task (gesture detection), the gesture detection function should be run as a background task. One of the background task implementations that have been tested is the python celery library, yet implementation of celery shows errors while running more than one celery task. Although the celery beat and celery worker did list out 2 of the celery tasks that have been defined, it will not execute the task and just kept prompting the message “Waking up now.” which shows in Figure 6. Figure 6. Sample screen of celery beat Thus, for current progression is to find another suitable approach to replace the celery library to run gesture detection as a background task. 101 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 5.0 CONCLUSIONS As one of the sub-projects of the TARUC Smart Campus initiative, this research project is aimed to provide Artificial Intelligence capabilities for the development of the dashboard employed in the i2hub, also known as the integrated innovation hub, located in the TARC cyber center which collaborates with works on initiatives such as Industry 4.0, Agriculture 4.0 and more. The dashboard that will be developed is planned to be available to cooperate with these initiatives by authorized assessment to the initiatives using the dashboard such as monitoring the robot motions from the video captured on the Industry 4.0 manufacturing site’s cameras and more. This project also envisaged the dashboard with manipulation functionalities to be performed using human gestures detection. Although there are lots of products with similar functionalities on the market, this research outcome aims to give a better solution for resolving the implementation issues faced in the real-world environment. Especially during this Covid-19 pandemic outbreak, the healthcare industry advocated making use of the interactive dashboard in better remote monitoring of the patients. This reduces unnecessary physical contact with patients or objects to control the dashboard which may be contaminated with viruses. Besides, the interactive dashboard could also be advantageous for the manufacturing industry during this pandemic. Due to the high rate of workers getting affected by the Coronavirus, some factories owned by Top Glove are required to close down (Teoh and Kalbana, 2020). This has sorely affected the production of the company. Then, it is nice for imposing the manufacturing works by applying machine automation and interactive dashboard in remote monitoring on the machines to take over the worker’s positions and continue the productions during the pandemic. (Chen and Lin, 2020.) Considering this project as one of the researches done in TARUC and the sub-project of the TARUC smart campuses project, this project is aimed to bring new findings and knowledge which are beneficial for the smart campus project and expect to aid for further studies with related topics. For instance, the video streaming protocols that have been studied in this research may be beneficial for further workings in the TARC Smart Campus’s IoT initiative to expand the infrastructure and merge with more systems and other related projects on the campus. Lastly, the zero-shot learning capabilities explored throughout the studies are expected to be valuable for the further related workings which require dealing with extensive quantities or insufficient size of data. REFERENCES Chakraborty, B.K., Sarma, D., Bhuyan, M.K. and MacDorman, K.F., 2017. “Review of constraints on vision-based gesture recognition for human-computer interaction.” IET Computer Vision, 12(1), pp.3-15. (Accessed: 20 Feb 2021). Chen, T. and Lin, C.W., 2020. Smart and automation technologies for ensuring the long-term operation of a factory amid the COVID-19 pandemic: an evolving fuzzy assessment approach. “The International Journal of Advanced Manufacturing Technology”, 111(11), pp.3545-3558. (Accessed: 3 April 2021). G. Sinha, R. Shahi, and M. Shankar, \"Human-Computer Interaction,\" 2010 3rd International Conference on Emerging Trends in Engineering and Technology, 2010, pp. 1-4, DOI: 10.1109/ICETET.2010.85. (Accessed: 09 Sep 2021). 102 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Gokul, V., Balakrishnan, G.P., Dubnov, T. and Dubnov, S., 2019, March. “Semantic Interaction with Human Motion Using Query-Based Recombinant Video Synthesis”. In 2019 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR) (pp. 379-382). IEEE. (Accessed: 20 Feb 2021). Guo, X., Xu, W., Tang, W.Q. and Wen, C., 2019, October. “Research on Optimization of static gesture recognition based on convolution Neural Network”. In 2019 4th International Conference on Mechanical, Control and Computer Engineering (ICM CCE) (pp. 398-3982). IEEE. (Accessed: 20 March 2021). Hikvision, 2019, “Network Camera User Manual UD15501B”, Hangzhou Hikvision Digital Technology Co., Ltd.: Author (Accessed: 27 March 2021). Köpüklü, O., Gunduz, A., Kose, N. and Rigoll, G., 2019, May. “Real-time hand gesture detection and classification using convolutional neural networks”. In 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019) (pp. 1- IEEE. (Accessed: 13 March 2021). Materzynska, J., Berger, G., Bax, I. and Memisevic, R., 2019. “The jester dataset: A large- scale video dataset of human gestures.” In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (pp. 0-0). (Accessed: 28 March 2021). Miroslav P. (12 February 2018) Introduction [Review of the book Communication Protocol Engineering, by Miroslav P.] Taylor & Francis Group, https://www.taylorfrancis.com/books/mono/10.1201/9781315151243/communication- protocol-engineering-miroslav-popovic(Accessed: 09 Sep 2021). Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K. and Woo, W.C., 2015. “Convolutional LSTM network: A machine learning approach for precipitation nowcasting.” arXiv preprint arXiv:1506.04214. (Accessed: 27 March 2021). Teoh P.Y. and Kalbana P., 2020. 28 “Top Glove factories in Klang to close down temporarily to enable Covid-19 testing, New Straits Times”, November 23, Available at: https://www.nst.com.my/news/nation/2020/11/643635/28-top-glove-factories-klang- close down-temporarily-enable-covid-19 (Accessed: 3 April 2021). 103 of 225 ICDXA/2021/9 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 A REVIEW ON THE DEVELOPMENT OF DATASPACE CONNECTORS USING MICROSERVICES CROSS- COMPANY SECURED DATA EXCHANGE Sze-Kai Gan1*, Thein-Lai Wong1, Ching-Pang Goh1, Wah-Pheng Lee2 and Yee-Mei Lim3 1 Faculty of Computing and Information Technology, 2 Centre for Postgraduate Studies and Research, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia 3 Tech Innovation and Production, GMCM Sdn. Bhd., Kuala Lumpur, Malaysia *Corresponding author: [email protected] ABSTRACT Industry Revolutions 4.0 (IR4.0) involves digital transformation of businesses to deliver quality products and services supported by real time decision making. It does not only impact the manufacturing industry but involves everyone in any value chains, from producers to consumers in all industries. With the integration of Internet of Things (IoT) devices and business applications, data generated provide valuable information for business insights, hence referred as assets. Industrial Data Space Reference Architecture Model (IDS-RAM) aims to achieve a reliable exchange of data between organizations and platforms developed by different vendors, hence enabling seamless value chain integration. Data flow between organizations are usually crucial, high load and real time. Therefore, the data exchange architecture that involves the dataspace connectors that reside in every organization, must enable uninterrupted quality service, real time streaming and secured usage control of data with high volume, high velocity and high availability. To handle the data communication, whether they are stored in premises, on the cloud, or on edge, with different computing capacity and capabilities, all systems need a flexible, light and reliable connector architecture that can scale and optimize resources effectively based on demand. This paper reviews the existing researches that are related to the development of dataspace connectors in the past 10 years. Keywords: High Availability, Industry 4.0, Reactive, Microservice, Connector 1.0 INTRODUCTION The Industrial Revolution (IR) began in the 18th century has been changing the working conditions and lifestyles of humans tremendously. From steam-powered factories to mass production and eventually computer technology, the revolution increased human productivity while reducing the cost significantly. The emergence of IR 4.0 has given rise to the new wave of fundamental change in industry revolution(Ahuett-Garza and Kurfess, 2018). IR 4.0 represents the result of automation combined with information technology and operation technology (Lu, 2017), such as Industry Internet of Things (IIoT) and cloud computing to improve the accessibility to real-time data. IIoT uses the power of smart sensors, actuators, and real-time data analytics to improve the manufacturing process. The smart sensors installed at different stations are interconnected and controlled by computer software to gather huge amounts of data for real-time analysis. This allows the company to monitor, identify and amend the inefficiencies and problems in manufacturing processes without any 104 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 delay. IIoT has been widely used in the field of manufacturing, warehouse, smart utilities, smart city, etc., which required full automation and monitoring. Reference Architecture Model Industry 4.0 (RAMI 4.0)1(Federal Ministry for Economic Affairs and Energy, 2019) was published in 2015 by the German Electrical and Electronic Manufacturers' Association (ZVEI), to support the Industry 4.0 efforts. RAMI 4.0 emphasizes industrial production as the primary area of application as shown in Figure 1. According to DIN SPEC 91345 (2016), the Information Layer describes the data that is used, generated or modified by the technical functionality. This includes a runtime environment for pre- processing of events, consistent integration of different data, formal description of models and rules, etc. The Communication Layer describes Industry 4.0-compliant access to the information and functions of the connected assets. In other words, it manages the exchange of data between company i.e., which data is used, where it is used and when it is distributed. Figure 1. Reference Architecture Model for Industry 4.0 (Federal Ministry for Economic Affairs and Energy, 2019). In 20172, Industrial Data Space Reference Architecture Model (IDS-RAM) enables a reliable exchange of data between company with common rules to be implemented in the Communication Layer of RAMI 4.0. Data sovereignty is, thus, a central aspect for IDS. Besides specifying the rules to ensure the data sovereignty within data ecosystems, the IDS- RAM, published by International Data Spaces Association (IDSA) (Otto et al., 2019), consists of five layers: business layer, functional layer, process layer, information layer and system layer. The standards materialize in the IDS-RAM and DIN SPEC 27070:2020-03 (2020) define methods for secure data exchange between the various Industrial Data Space connectors. Figure 2 shows each connector is able to communicate with every other connector or component in the ecosystem of the Industrial Data Space, to allow any organization to communicate to the outside world real-time. 1 https://www.i-scoop.eu/industry-4-0/#The_Reference_Architectural_Model_Industrie_40_RAMI_40 2 https://www.i-scoop.eu/industry-4-0/industrial-data-space/ 105 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Figure 2. Data Exchange Architecture in IDS environment This paper examines the current research related to dataspace connector development and the issues of data security, high availability, low latency and scalable resources of cross-company data exchange between two trusted connected are not addressed in IDS-RAM. It is particularly crucial for the development of digital twin for timely distributed business operations in the Industry 4.0. Hence, this paper will review and propose the reactive microservices as an architecture model to overcome the challenges of using Data Exchange Architecture in the cross-company secured data exchange in the manufacturing environment. 2.0 LITERATURE REVIEW Detailed research related to trusted connector development over the recent decade are studied in this section. The research findings are categorized into three general areas which are service-oriented architecture (SOA) and microservices, orchestration and security. 2.1 Service-Oriented Architecture (SOA) and Microservices Microservices architecture (Fowler and Lewis, 2014) is a variant of the service-oriented architecture (SOA) (Xiao, Wijegunaratne and Qiang, 2017), which is built from a collection of services. Each individual service contains very minimal functions and runs on its own process autonomously. In contrast to the monolithic architecture which comprises all modules into a single application, the collection of the microservices is loosely coupled; hence it provides the flexibility to modify and upgrade the services without interrupting other modules. In addition, microservices can achieve autonomy, isolation, and resilience with the help of external services, for instance load balancer. A recent study done by Santana, Alencar and Prazeres (2018) showed microservices architecture started being used in IoT applications to improve the performances of large-scale systems. Petrasch (2017) studied the microservices architecture with Enterprise Integration Patterns (EIP)(G. Hohpe, 2003) for inter-service communication. They demonstrated how domain driven design, model-driven and pattern- based approach for composing microservices to reduce the complexity of microservices architecture. In 2019, Bigheti, Fernandes and Godoy proposed a Microservice-Oriented Architecture (MOA) for Control as a Service (CAAS). The authors simplified the MOA deployment, and 106 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 provided a flexible, interoperable, and distributed architecture. It is a high-level design of the microservices as enhancement for ISA95 architecture, but they did not further address the drawbacks on this architecture and their solution. Dinh-Tuan, Beierle and Garzon (2019) implemented the microservices architecture in industrial manufacturing environments. They designed a decentralized architecture with the microservices based for industrial data analytic by developing the prototype, analyzed, and evaluated to provide further practical insights. Reactive microservice further enhance the standard microservice to have more isolation and autonomy with its consensus protocol and asynchronous non-blocking messaging architecture3. Later in 2021, Santana et al. improved the limitation of high availability by utilizing the reactive and asynchronous programming paradigm in his research. The result showed that the reactive system improved only the availability of the services of distributed IIOT networks in agriculture 4.0. However, there is still lack of research using reactive microservices to improve the requirements of timely data exchange in the distributed manufacturing environment To enhance and manage the complex architecture defined in IDS-RAM, each of the module shall work independently and autonomy with microservice architecture. Modules could be delegated to specialized team to develop and deliver. 2.2 Service Orchestration Service orchestration is the arrangement, configuration and management of the system to ensure the system operates within its ecosystem with the expected behavior.4 Girbea et al. (2014) designed a novel architecture in industrial automation based on the SOA. The new architecture was able to compute optimal production plans and executed the plans automatically. Besides, the development and the maintenance were also flexible and reusable. This enhanced the seamless transition from the current practice. Their paper focused on data monitoring in manufacturing processes without considering the data exchange between organizations. Theorin et al. (2017) developed an event-driven architecture for IR4.0, which was Line Information System Architecture (LISA). This architecture enabled the data exchange between organizations in more efficient and flexible ways. It also allowed data integration and utilization by using SOA and Enterprise Services Bus (ESB). In the Manufacturing as a Service (Maas) domain, Landolfi et al. (2019) developed an ecosystem that acted as a virtual marketplace. This virtual marketplace promoted the virtual and physical assets so that the production demand was able to be achieved optimality. For the next evolution of the architecture, the authors suggested using FIWARE Generic Enablers (GE)(Alonso et al., 2018) and an extensive testing of their functionalities. FIWARE5 offers an open standard platform and a set of standard APIs to support IoT and smart applications development in various domains. A FIREWARE GE is a set of general-purpose platform functions available through APIs. Trunzer et al. (2019) discussed some system RAMI 4.0, American Industrial Internet Reference Architecture (IIRA) (Lin et al., 2017), Line Information System Architecture (Theorin et al., 2017), SORRADES architecture (Karnouskos, Bangemann and Diedrich, 2009), etc. to address several issues in cross-enterprise data sharing, service orchestration and 3 https://dzone.com/articles/diving-into-reactive-microservices 4 https://cloudify.co/blog/why-service-orchestration-matters/ 5 https://www.fiware.org 107 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 real-time capabilities. These generic architectures bridged up the gap between reference and use-case specific architectures. Microservices or SOA are highly flexible and dynamic and require more complex system architecture. Studies showed the needs of service orchestration to ensure the interconnectivity of the services and functions are performed as good as monolith application. Some modern orchestration tools such as Kubernetes, Chef, Ansible could help to manage those complex design. 2.3 Trusted Data Security Security has always been prioritized in data exchange as the data is sensitive and it is the most important asset in an organization. A reliable security policy and algorithm is able to safeguard the data integrity, authority, and privacy especially in the manufacturing industry where operational and business data is proprietary and confidential. A study from Li, Xuan and Wen (2011) on the general architecture of trusted security system based on IoT materialized the trusted user module, trusted perception module, trusted terminal module, trusted network and trusted agent module to eliminate the various security threats in the application of IoT. Later in 2017, Teslya and Ryabchikov (2017) implemented the blockchain in IoT to enhance the security level of the IoT system. The authors designed an architecture that combined Smart-M3 information sharing platform, blockchain platform and smart contracts to process and store the information between smart space components. Nevertheless, this architecture has several limitations which are (1) lack of mechanisms to establish authorship, durability and unchangeability of information, (2) control over the exchange of resources in production, and (3) an integrated mechanism for reaching consensus among participants. (Brost et al., 2018) developed a comprehensive security architecture which included Service Manager, Connection Manager, Routing Manager, Audit Logging, Dataflow control, Access control and Usage Control, Security Management as services bundled in OSGI framework (OSGi Alliance, 2018). The major drawback of the OSGI framework is it required all the services included in the single OSGI container for all the services to interact and run as a monolith application. Munoz-Arcentales et al. (2019, 2020) enabled the access and usage control in data-sharing ecosystems among multiple organizations in the food industry using the FIWARE European opensource platform. The Policy Enforcement Point (PEP) intercepted the incoming request from the Data Consumer to the Data Provider’s infrastructure to enforce access control on resources. In the Healthcare cyber physical system, Xu et al. (2020) proposed a privacy-preserving data integrity verification model by using lightweight streaming authenticated data structures. The authors discussed the design idea, architecture, formal definition, security definition and communication protocols in the model. In the Streaming Authenticated Data Structures (SADS), author provide six probability polynomial algorithms in to the model and leveraging the computing of the fully homomorphic encryption (FHMT) which shifts almost all the computation tasks to the server and little overhead for the client. Otto and Jarke (2019) designed a multi-sided data platform (MSP) based on IDS for secure and trusted data exchange but they only conducted a preliminary study and design on MSP’s lifecycle Data security and sovereignty are one of the key features for IDS-RAM, thus the data connector need to ensure the data exchange between manufactures are secured. 108 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 3.0 DISCUSSION AND MODELS Although many researchers have developed different architectures to improve the manufacturing process and data exchanges for an organization, some other issues such as availability, fault tolerance and services discovery are still remained unsolved. Baboi, Iftene and Gîfu (2019) showed that microservices architecture can improve fault tolerance and scalability using various technologies from .Net to Java. Their studies showed how flexible and dynamic the microservices are, but the fault tolerance capability is built in a very basic way, which is to remove the server from the list if the server does not respond in the timely manner. Microservices architecture has been widely used in many fields to improve modularity, allow integration with different systems, platforms, and legacies. However, there are some other issues to be addressed. For instance, how to design an effective dynamic service allocation algorithm, to optimize the system’s performance in runtime by balancing between computing power and latency. Besides, it is also important to have effective messaging techniques that could further improve the latency issue of microservices. Some other specific challenges when designing data centric multi-sided data platform (MSP), for example security, performance, data exchange optimisation, should be addressed too. Figure 3. Zookeeper architecture A centralized service discovery such as Zookeeper6 can be developed to store the metadata of the services, service registration and discovery as shown in figure 3. Zookeeper ensures data consistency based on CAP theorem (Fox and Brewer, 1999). With this centralized approach, Zookeeper is able to provide actual status of the services and manage the services within the cluster. However, this architecture is highly dependent on the Zookeeper’s performance. If the Zookeeper is down or delay in the metadata exchange when a new Zookeeper is added to the cluster, those services that are under management will not be able to communicate with each other. Besides, it has always been a challenge to design a secure microservices ecosystem that ensures data integrity and security as most of the services are communicated remotely. In addition, a trusted gateway needs to be built to verify and route the authorized request from the external world to the respective microservices. Therefore, a secure, responsive, high resilience and elasticity are very important criteria to build an ideal dataspace ecosystem. To overcome the above-mentioned challenges, reactive systems, introduced in the Reactive Manifesto (Bonér et al., 2014), enables consensus protocol, such as gossip protocol, to manage the distributed systems in a decentralized architecture to increase the isolation and autonomy of the system. Cluster modules manage its own cluster membership, load balancing and node partitioning, and achieve high availability based on CAP theorem. The combination 6 https://www.ibm.com/analytics/hadoop/zookeeper 109 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 of the simplicity and flexibility of the microservices architecture with reactive system principles, also known as Reactive Microservices Architecture, may help to solve the challenges. The implementation of the reactive microservices should be done based on the Reactive Manifesto four principles: responsive, resilient, elastic and message-driven. The role of the service discovery like zookeeper, consul will act as service lookup to allow cluster modules to be access by external services. 4.0 CONCLUSIONS Based on the findings of literature review in the four main areas, namely Service-Oriented Architecture (SOA) and microservices, service orchestration, systems interoperability and trusted security. The SOA and the microservices development have the following advantages: (1) more flexible, (2) could decouple the applications, (3) enhance the maintainability, (4) reduce complexity of the software architecture, (5) increase the scalability and performance. However, the studies showed that the architecture increases the complexity to manage and monitor the huge numbers of services running on the premises in the system. The main challenges which are not explored in the past researches is how to manage the multiple clusters of the services in more efficient way without losing any single data. Reactive microservices architecture is believed to achieve a fault tolerant and fast data processing system, enabling value-chain integration and real-time data exchange as recommended in RAMI 4.0 and IDS-RAM. For interconnectivity and handling of unlimited stream of data from various sources such as IoT, on-premise servers, cloud, edge devices between data providers and data consumers, traditional HTTP or RESTFul protocol usually having difficulty to handle the huge load of the data, since each request needs a complete response result before it releases the connection and thread. The device or application that accepts huge data via the HTTP protocol required to keep the data in memory until the data was completely sent out. Therefore, it usually requires a large amount of memory with traditional monolith application architectural design. To address these issues, an asynchronized data streaming technology shall process the data in small chunk and to have the backpressure capability to control the amount of data transferred from the source to prevent data overflow.7 The future work is to develop a dataspace connector with reactive microservice to offer a system with seamless data connectivity and high availability for data exchange between organizations in the industrial dataspace. REFERENCES Ahuett-Garza, H. and Kurfess, T. (2018) ‘A brief discussion on the trends of habilitating technologies for Industry 4.0 and Smart manufacturing’, Manufacturing Letters, 15, pp. 60–63. doi: 10.1016/j.mfglet.2018.02.011. Alonso, Á. et al. (2018) ‘Industrial Data Space Architecture Implementation Using FIWARE’, Sensors, 18(7), p. 2226. doi: 10.3390/s18072226. Baboi, M., Iftene, A. and Gîfu, D. (2019) ‘Dynamic microservices to create scalable and fault tolerance architecture’, Procedia Computer Science, 159, pp. 1035–1044. doi: 10.1016/j.procs.2019.09.271. Bigheti, J. A., Fernandes, M. M. and Godoy, E. D. P. (2019) ‘Control as a Service: A Microservice Approach to Industry 4.0’, 2019 IEEE International Workshop on Metrology for Industry 4.0 and IoT, MetroInd 4.0 and IoT 2019 - Proceedings, pp. 438– 443. doi: 10.1109/METROI4.2019.8792918. 7 https://nordicapis.com/rest-vs-streaming-apis-how-they-differ/ 110 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Bonér, J. et al. (2014) ‘The Reactive Manifesto (Version 2.0)’, Reactivemanifesto.Org, 2(16 September 2014), p. pp.1-2. Available at: http://www.reactivemanifesto.org. Brost, G. S. et al. (2018) ‘An ecosystem and IoT device architecture for building trust in the industrial data space’, CPSS 2018 - Proceedings of the 4th ACM Workshop on Cyber- Physical System Security, Co-located with ASIA CCS 2018, pp. 39–50. doi: 10.1145/3198458.3198459. ‘DIN SPEC 27070:2020-03’ (2020). doi: 10.31030/3139499. ‘DIN SPEC 91345’ (2016) in. Berlin, Heidelberg: Springer Berlin Heidelberg. doi: 10.31030/2436156. Dinh-Tuan, H., Beierle, F. and Garzon, S. R. (2019) ‘MAIA: A microservices-based architecture for industrial data analytics’, Proceedings - 2019 IEEE International Conference on Industrial Cyber Physical Systems, ICPS 2019, pp. 23–30. doi: 10.1109/ICPHYS.2019.8780345. Federal Ministry for Economic Affairs and Energy (2019) ‘Plattform Industrie 4.0 - RAMI4.0 – a reference framework for digitalisation’, Plattform Industrie 4.0. Fowler, M. and Lewis, J. (2014) Microservices. Available at: https://martinfowler.com/articles/microservices.html. Fox, A. and Brewer, E. A. (1999) ‘Harvest, yield, and scalable tolerant systems’, Proceedings of the Workshop on Hot Topics in Operating Systems - HOTOS, pp. 174–178. doi: 10.1109/hotos.1999.798396. G. Hohpe, B. W. (2003) Enterprise Integration Patterns: Designing, Building, and Deploying Messaging Solutions. Addison-Wesley. Girbea, A. et al. (2014) ‘Design and implementation of a service-oriented architecture for the optimization of industrial applications’, IEEE Transactions on Industrial Informatics, 10(1), pp. 185–196. doi: 10.1109/TII.2013.2253112. Karnouskos, S., Bangemann, T. and Diedrich, C. (2009) ‘Integration of legacy devices in the future SOA-based factory’, IFAC Proceedings Volumes (IFAC-PapersOnline), 13(PART 1), pp. 2113–2118. doi: 10.3182/20090603-3-RU-2001.0487. Landolfi, G. et al. (2019) ‘A MaaS platform architecture supporting data sovereignty in sustainability assessment of manufacturing systems’, Procedia Manufacturing, 38(Faim 2019), pp. 548–555. doi: 10.1016/j.promfg.2020.01.069. Li, X., Xuan, Z. and Wen, L. (2011) ‘Research on the architecture of trusted security system based on the internet of things’, Proceedings - 4th International Conference on Intelligent Computation Technology and Automation, ICICTA 2011, 2, pp. 1172–1175. doi: 10.1109/ICICTA.2011.578. Lin, S.-W. et al. (2017) ‘The Industrial Internet of Things Volume G1 : Reference Architecture’, Industrial Internet Consortium White Paper, Version 1., p. 58 Seiten. Lu, Y. (2017) ‘Industry 4.0: A survey on technologies, applications and open research issues’, Journal of Industrial Information Integration, 6, pp. 1–10. doi: 10.1016/j.jii.2017.04.005. Munoz-Arcentales, A. et al. (2020) ‘Data usage and access control in industrial data spaces: Implementation using FIWARE’, Sustainability (Switzerland), 12(9). doi: 10.3390/su12093885. OSGi Alliance (2018). Available at: https://www.osgi.org/developer/architecture/. Otto, B. et al. (2019) ‘International Data Space Reference Architecture Model Version 3.0’, (April), pp. 1–118. Available at: https://www.internationaldataspaces.org/wp- content/uploads/2019/03/IDS-Reference-Architecture-Model-3.0.pdf. Otto, B. and Jarke, M. (2019) ‘Designing a multi-sided data platform: findings from the International Data Spaces case’, Electronic Markets, 29(4), pp. 561–580. doi: 10.1007/s12525-019-00362-x. Petrasch, R. (2017) ‘Model-based engineering for microservice architectures using Enterprise 111 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Integration Patterns for inter-service communication’, Proceedings of the 2017 14th International Joint Conference on Computer Science and Software Engineering, JCSSE 2017, pp. 2–5. doi: 10.1109/JCSSE.2017.8025912. Santana, C. et al. (2021) ‘Increasing the availability of IoT applications with reactive microservices’, Service Oriented Computing and Applications, 15(2), pp. 109–126. doi: 10.1007/s11761-020-00308-8. Santana, C., Alencar, B. and Prazeres, C. (2018) ‘Microservices: A mapping study for internet of things solutions’, NCA 2018 - 2018 IEEE 17th International Symposium on Network Computing and Applications. doi: 10.1109/NCA.2018.8548331. Teslya, N. and Ryabchikov, I. (2017) ‘Blockchain-based platform architecture for industrial IoT’, in 2017 21st Conference of Open Innovations Association (FRUCT). IEEE, pp. 321–329. doi: 10.23919/FRUCT.2017.8250199. Theorin, A. et al. (2017) ‘An event-driven manufacturing information system architecture for Industry 4.0’, International Journal of Production Research, 55(5), pp. 1297–1311. doi: 10.1080/00207543.2016.1201604. Trunzer, E. et al. (2019) ‘System architectures for Industrie 4.0 applications: Derivation of a generic architecture proposal’, Production Engineering, 13(3–4), pp. 247–257. doi: 10.1007/s11740-019-00902-6. Xiao, Z., Wijegunaratne, I. and Qiang, X. (2017) ‘Reflections on SOA and Microservices’, Proceedings - 4th International Conference on Enterprise Systems: Advances in Enterprise Systems, ES 2016, pp. 60–67. doi: 10.1109/ES.2016.14. Xu, J. et al. (2020) ‘Privacy-preserving data integrity verification by using lightweight streaming authenticated data structures for healthcare cyber–physical system’, Future Generation Computer Systems, 108, pp. 1287–1296. doi: 10.1016/j.future.2018.04.018. 112 of 225 ICDXA/2021/10 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 TEXT SUMMARIZATION ON AMAZON FOOD REVIEWS USING TEXTRANK Yuen Kei Khor1*, Chi Wee Tan1 and Tong Ming Lim2 1 Faculty of Computing and Information Technology, 2 Centre For Business Incubation And Entrepreneurial Ventures, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia *Corresponding author: [email protected] ABSTRACT Text summarization is a technique to create a summary by shortening the length of text but keep the key information. There are two main approaches to summarize the text which are abstractive summarization and extractive summarization. This study is aimed to extract the most important top 5 reviews which can summarize the overall reviews of certain product in Amazon fine food reviews. TextRank algorithm which is one of the extractive summarization approaches is used to perform text summarization automatically. GloVe pre-trained word embedding model with 100 dimensions is used to map each word from the reviews to vector representation. Besides, PageRank algorithm is applied to compute the sentence rankings scores to determine how important and relevant of the sentences can be the representatives of summary. Top 5 reviews with the highest sentence ranking scores are extracted to be the summary and further discussed the customer perception on the product based on the summary generated. The final summary shows that Amazon customer reviews tend to positive for certain food brand. Keywords: Text Summarization, Extractive Summarization, TextRank List of notations n is the number of individual sentences 1.0 INTRODUCTION With the advent of technology, social media platforms and websites have become a sharing platform for the public to share their experiences and thoughts towards products, services or latest news freely. This massive amount of text contributes useful information but it is not efficient for human to read entirely and create a summary so text summarization comes into rescue. Maybury (1999) defined text summarization as the process of distilling the most important information from a source (or sources) to produce an abridged version for a particular user (or users) and task (or tasks). In other words, text summarization produces a summary from one or more plain text while preserving the key information. There are two approaches used for automatic text summarization, abstractive summarization and extractive summarization. Abstractive summarization used advanced natural language techniques such as deep learning approach to generate a totally new and shorter text that preserve the important information from the original text document. In contrast, extractive summarization is the extraction of a subset of important sentences from the original text. 113 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 This study is motivated in order to replace human power to summarize a sizeable text into a few sentences. Other than time consuming issues, the quality of summaries are highly depends on human knowledge and language ability level. Sometimes they may misunderstand the content meaning of text document. In this study, we applied extractive summarization approach to summarize text in an automatic way. Besides, it often outperform than abstractive summarization. Allahyari et al. (2017) stated that abstractive summarization need to deal with problems such as semantic representation, inference and natural language generation which is a challenging task for sentence extraction. This objective of this study is to extract top 5 reviews with the highest sentence ranking scores which can summarize the overall reviews of certain product from Amazon fine food reviews. The top sentence represents the highest chances of the topic discussed from customers in overall product reviews of that particular product. With top ranked sentences from the summary generation, we can roughly know the opinions or comments of customers towards that particular food. 2.0 LITERATURE REVIEW In this project, we will focus on the extractive summarization approach and discuss the main approaches which are widely used in research work. There are three distinguished independent tasks to perform extractive summarization: construct an intermediate representation of the document, scoring sentences based on the representation, select few most important sentences to create a summary according to the scoring. There are few common approaches widely used in text summarization that will further discussed below. In the early research on extractive summarization, researchers use features from the sentences such as their position in the text, word frequency, or key phrases indicating the importance of the sentences (Erkan & Radey, 2004). Term Frequency-Inverse Document Frequency (TF-IDF) method is used to determine how important the word in a document is. TF-IDF values increase when the number of times of a word appears is increased in a document. This method works in the weighted term-frequency and inverse sentence frequency. Sentence frequency is the total number of sentences containing that term in the document. The sentence vectors will be scored by similarity and the sentences with highest similarity scores will be picked as a part of summary (Saranyamol & Sindhu, 2014). Christian et al. (2017) proposed an automatic text summarizer which implement TF-IDF to extract three to five sentences with the highest TF-IDF scores to be the final summary, where the number of sentences are decided by users. Results shows the proposed system produces 67% accuracy by comparing with other online automatic summarizer. Machine learning can be applied for text summarization if the dataset or documents has the respective extractive summary. This is because machine learning required a large amount of labelled data to train the model. The model will learn the patterns by identifying those relevant features values that are correlated with the labelled data. Therefore, the feature extraction process is very crucial in order to improve the accuracy of the summarization. Having more training data will also increase the accuracy as the model can learn more patterns so that sentences can be produced an extractive summary when new documents are given to the model. Neto et al. (2002) present a text summarizer using two well-known algorithm, Naive Bayes and C4.5 decision tree algorithm with a set of features that are statistics-oriented and linguistic-oriented. Performance of these two algorithms are compared with two baseline method with two sets of experiment by employing automatically-produced extractive summaries and manually-produced summaries. Results shows that Naïve Bayes outperform all the summarizers. Deep learning approach also quite common in automatic text summarization (PadmaPriya, 2014; Day & Chen, 2018; Patel et al., 2018). 114 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 There are two elements that should be known when we talk about the graph-based approach which are nodes and edges. Nodes are the sentences and edges are the similarity between two sentences. Two sentences are connected with an edge if they share some common words. If the nodes have a number of edges connected to it, then they are considered as important sentences and carry high preference to be included in the summary. TextRank is a well-known graph-based approach which applied on text summarization. TextRank is inspired by PageRank (Brin & Page, 1998) which is implemented by Google, where TextRank is used to rank sentences while the latter rank web pages. In PageRank, web pages are linked with other important web pages so they will have high PageRank value. PageRank can be applied in text summarization to select the most important sentences from original text document. In the study of Mallick et al. (2019), they proposed a modified PageRank algorithm where assume that the important sentences are linked (similar) to other important sentences in the text document. Li and Zhao (2016) also proposed a TextRank algorithm by exploiting Wikipedia for short keywords extraction. Their findings show TextRank model constructed based on Wikipedia as external knowledge works better than traditional TextRank which use TF-IDF. 3.0 METHODOLOGY AND FRAMEWORK This section shows the framework of TextRank algorithm in Figure 1 and a detailed description for each steps are discussed. Figure 1. TextRank flowchart of Amazon fine food reviews summarization. 115 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 3.1 Experiment 3.1.1 Dataset This project used Amazon fine food reviews which is available from Kaggle to perform text summarization. The reviews are collected from the fine foods of Amazon. This dataset contains 568,454 reviews to 74,258 of products which are collected from October 1999 to October 2012. 3.1.2 Data Extraction The dataset includes products and users’ information, rating, number of users who found the review helpful, reviews and summary of reviews. In this experiment, reviews and summary of reviews are being extracted for the most commented product in the dataset. It is a cookie brand but brand name will not be disclosed due to privacy reason. The dataset is reduced to 910 after extracting the related reviews and dropping duplicates. 3.1.3 Sentence Segmentation A review may consist of few sentences so it is necessary to segment the reviews into an individual sentence in order to proceed for further steps. Reviews are segmented when ends with question marks, exclamation mark and full stop. A total of 910 reviews are segmented to 3661 individual sentences. 3.1.4 Text Preprocessing Textual reviews collected from Amazon are noisy and greatly impact on the performance of text summarizing. Noises such as HTML tags, punctuations and stop words are removed and all letters converted to lowercase. 3.1.5 Extract Word Vectors Machine could not understand the semantic and syntactic similarity between words in a document. Therefore, techniques are applied to convert each word to real-valued vector which called word embedding. Word embedding is a real-valued vector that represents an individual word in a vector space. Figure 2 shows an example graph of word embedding, each word will map to a word embedding. Words that tend to occur in same contexts will close to each other in the vector space (McDonald & Ramscar, 2001). Based on the graph, ‘cookie’ and ‘biscuit’ are close to each other so it can be deduced that these two terms have related meaning. In this experiment, Global Vectors (GloVe) is used to convert each word in the sentences to word vectors. GloVe (Pennington et al., 2014) is an unsupervised learning algorithm for obtaining vector representations for words. It is trained on the nonzero element of global word-to-word co-occurrence by estimating how frequently words co-occur with one another in a given corpus. Pre-trained word vectors with 100 dimensional of 400k words computed on 2014 dump of English Wikipedia is used to create vectors for sentences. Each word will have 100 vectors in the 100 dimensional pre-trained word vectors. The trained word vector is available at (https://nlp.stanford.edu/projects/glove/). Next, 100 vectors of each word will be fetched and calculate the total vectors of each word in the sentence. The sum of vectors will be divided by the total number of words in a sentence and this will be the final vector. 116 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Figure 2. Word Embedding. 3.1.6 Cosine Similarity Score Even though TF-IDF is commonly used in text summarization to calculate the relevance and importance of sentences but we will used cosine similarity to find the similarity between sentences. This is because TF-IDF are too long and sparse (having many zero values as sentences may not share the same words). Despite there are no common words occurred in two sentences but this does not represents the sentences have no related meaning (Han et al. 2012). In contrast, cosine similarity measure only focus on common words between sentences and measure the similarity between sentences. Therefore, a zero similarity matrix (n*n) is created. Cosine similarity is used to compute the similarity scores between sentences vectors and assigned to the matrix. This matrix is called as similarity matrix. If the score is 0, there is no relationship between these two sentences. 3.1.7 PageRank Algorithm Next, similarity matrix is converted into graph which consists of two elements, nodes and edges. Nodes represent the sentence and edges refers to the similarity scores between the sentences. With the graph, PageRank algorithm is used to compute the sentence rankings scores. The scores are used to determine how important and relevant the sentences are to form the summary. 3.1.8 Extract top-ranked sentences Sentence ranking scores are sorted in descending order to get the top-ranked sentences which are the representatives of the summary of this cookie reviews. In this project, top 5 highest sentence ranking scores are extracted to form the summary of cookie reviews. 4.0 RESULT Table 1 shows the top 5 sentences with the highest sentence ranking scores which can be concatenated to be the summary of cookies reviews. By looking through the sentences, we can see that the customers who gave comments for the first and third sentences got a cookie sample from Influenster, a product discovery and review platform, and they love it very much due to the softness or freshness. In contrast, the second and fifth comments show that customers do not like the cookies as they are not fresh, dry and not crumbled. Meanwhile, the third comment also love the taste and softness. We can deduce that customers love the oatmeal flavour of this cookie brand. In an overall view, this cookie brand received compliments which is related to the taste, softness and freshness but it also dislikes by customers as the cookies are too dry, not crumble and not fresh enough. 117 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Table 1. Top 5 highest-ranked sentences. Ranking Summary 1 I GOT TO TRY THIS QUAKER SOFT BAKED OATMEAL 2 COOKIE THROUGH THE GOOD FOLKS FROM INFLUENSTER AFTER RECEIVING THEIR 2012 MOMVOX BOX, AND I MUST 3 SAY I LOVE IT, FIRST OF ALL OATMEAL COOKIES ARE MY 4 FAVORITE, SO THERE WASNT ANY DISAPPOINTMENT 5 THERE, THE COOKIE RETAIN ITS SOFTNESS/FRESHNESS OFTER BEING OPENED BY ME FOR A WEEK NOW, AND THAT WAS GOOD, PLUS IT TASTE GREAT SO THUMBS UP Maybe it was the baking process? These cookies, although individually packed (so good for school lunches), came out a bit dry and crumbly. Sure, maybe I am just a messy eater but a soft baked cookie just not crumble as much as the cookies I got crumbled. Maybe if you get them at the supermarket they would be less dry. Maybe if is just a general problem with the way they are produced. i recived a free sample from Influenster and let me tell you it was so good and soft it crumbles up right in your mouth and its a big cookie my daughter also loved it i would difenitly recommend buying it if you like oatmeal and raisins yummy great cookie just like my momma makes this is deffinietely a second best of course after my moms cooking love how soft and chewy they are a must buy I love soft baked cookies, but I find that whenever i try to buy ones that are already made, they don't taste fresh. 4.1 Limitations The limitation of this experiment is the computation time of the TextRank algorithm is very long. It takes few hours to complete the process of computing similarity matrix and sentence ranking scores for about 3600 of sentences. Computation time increase if the number of sentences extracted to perform summarization increases. This is because the similarity matrix size will increase and spend more time to compute the similarity scores between the sentences. Even though the summary is formed by top 5 original reviews with the highest-ranking score from the fine food reviews but it is still considered long to read. Sometimes the reviews can be wordy and difficult to read in a single glance. Researchers may not satisfy the result of summarization as it is still not a complete summarized version of the reviews. The accuracy of the summary of food reviews should take into consideration. This is due to the reason that only 100 vectors are used for each word to compute the vector representation of the sentence. 4.2 Future work The compute time can be reduced by generating more stopwords without using the default stopword list from Natural Language Toolkit (NLTK) library. With the list of stopwords, we can remove more words in the sentences during text preprocessing. Therefore, removing more stopwords from the sentences can greatly reduce the number of words to create vectors 118 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 for each of them. Since the number of words is decreasing, the compute time of calculating the average vectors of each word and sum of vectors for each sentence will be reduced too. Besides, we need to explore other techniques to summarize each review into few words and apply the algorithm used in this experiment to get a shorter summary. To increase the accuracy of summarization, we can fetch more vectors for each word in the sentence. GloVe word embedding consists of pre-trained word vector models with 50 dimension, 100 dimension, 200 dimension, 300 dimension for each words. We can extract word vectors from 200 dimension or 300 dimension pre-trained word vector model to create the vector representations for the sentences. However, higher dimension of pre-trained word embedding model used shall improve the accuracy but required longer computation time. We will explore more in order to take account of the speed and accuracy for this project. 5.0 CONCLUSIONS In this study, automatic text summarization is done on a cookie brand Amazon review using TextRank algorithm. Top 5 reviews with the highest sentence ranking score are extracted and formed to be the summary. Based on the summary, we can conclude that this cookie brand taste good, fresh and soft but some reviewers think it is dry, not fresh and crumble enough. Other than that, oatmeal cookie could be the most-favoured product to customer as high compliments received from them. The overall summary is tend to positive but still need improvement on the moistness and make it more crumble. With this summary, we are able to understand customers’ perception on the food product without read it entirely and write a summary. This study has shown that long computation time and unsatisfactory summary length for summarizing the reviews and improvement needs to be done in the future. 6.0 ACKNOWLEDGMENTS Authors would like to express my great appreciation to Tunku Abdul Rahman University College for the financial and resources support to implement this project. REFERENCES Allahyari M, Pouriyeh S, Assefi M et al. (2017) Text summarization techniques: a brief survey. International Journal of Advanced Computer Science and Applications (IJACSA) 8(10), http://dx.doi.org/10.14569/IJACSA.2017.081052. Brin S and Page L (1998) The anatomy of a large-scale hypertextual web search engine. Computer networks and ISDN systems 30(1-7): 107-117, http://dx.doi.org/10.3844/jcssp.2014.1.9. Christian H, Agus MP and Suhartono D (2016) Single document automatic text summarization using term frequency-inverse document frequency (TF-IDF). ComTech: Computer, Mathematics and Engineering Applications 7(4): 285-294, http://dx.doi.org/10.21512/comtech.v7i4.3746. Day MY and Chen CY (2018) Artificial intelligence for automatic text summarization. In 2018 IEEE International Conference on Information Reuse and Integration (IRI), pp. 478-484, http://dx.doi.org/10.1109/IRI.2018.00076. Erkan G and Radev DR (2004) Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of artificial intelligence research 22(1): 457-479. Han JW, Kamber M and Pei J (2012) Getting to know your data. In Data mining (Third Edition), pp. 39-82, https://doi.org/10.1016/B978-0-12-381479-1.00002-2. Li W and Zhao J (2016) TextRank algorithm by exploiting Wikipedia for short text keywords extraction. In 2016 3rd International Conference on Information Science and Control Engineering (ICISCE), pp. 683-686, https://doi.org/10.1109/ICISCE.2016.151. 119 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Maybury M (1999) Advances in automatic text summarization. MIT press. McDonald S and Ramscar M (2001) Testing the distributioanl hypothesis: The influence of context on judgements of semantic similarity. In Proceedings of the Annual Meeting of the Cognitive Science Society 23(23). Neto JL, Freitas A and Kaestner CA (2002) Automatic text summarization using a machine learning approach. In Brazilian symposium on artificial intelligence, Berlin, Heidelberg, vol. 2507, pp. 205-215, http://dx.doi.org/10.1007/3-540-36127-8_20. PadmaPriya G (2014) An approach for text summarization using deep learning algorithm. International journal of trends in computer science 10(1): 1-9, http://dx.doi.org/10.3844/jcssp.2014.1.9. Patel M, Chokshi A, Vyas S and Maurya K (2018) Machine Learning Approach for Automatic Text Summarization Using Neural Networks. International Journal of Advanced Research in Computer and Communication Engineering 7(1), http://dx.doi.org/10.17148/IJARCCE.2018.7133. Saranyamol CS and Sindhu L (2014) A survey on automatic text summarization. International Journal of Computer Science and Information Technologies, 5(6): 7889-7893. Pennington J, Socher R and Manning CD (2014) Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532-1543. 120 of 225 ICDXA/2021/11 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 A FLOWER RECOGNITION SYSTEM USING DEEP NEURAL NETWORK COUPLED WITH VISUAL GEOMETRY GROUP 19 ARCHITECTURE Zi Yuan Ong1*, Kah Kien Chye1 , Huay Wen Kang1 and Chi Wee Tan1 1 Faculty of Computing And Information Technology, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia *Corresponding author: [email protected] ABSTRACT Computer vision is one of the basic features to streamline processes like robotic process automation and digital asset management. Computer vision has come a long way in terms of its capabilities and what it can provide and do for different industries. Object detection and image detection are just some of the few applications provided by computer vision. However, this field is still relatively young and prone to challenges. The first is the lack of well- annotated images to train the algorithms to perform optimally, and the second being lack of accuracy when applied to real-world images different from the ones from the training dataset. As such, this paper aims to fine-tune pre-trained machine learning models, which are ResNet50 and VGG19 as well as training a new SqueezeNet inspired model from scratch to create a flower recognition model that can process and remember large amounts of flower species data. In conclusion, VGG19 was found to perform the best on both the 5 Categories and Flower-102 dataset, with an accuracy of 88% and 84% respectively. Keywords: VGG19, Transfer Learning, Deep Learning, Flower Recognition, Neural Network 1.0 INTRODUCTION There are approximately 369,000 named flowering plant species in the world (Liu et al., 2016). In general, experienced plant taxonomists can identify plants based on the features of flowers such as sepals, petals, stamens, and carpels. However, most people find it tough to determine these flowers apart. Additionally, someone may be confused with similar flower species. This is where object recognition comes in, as it is able to understand and analyze images effortlessly and instantaneously. Therefore, the main objective of this project is to develop a flower recognition model that can correctly identify the class of flowers. The flower recognition model will analyse the image and to identify whether an input image contains a certain type of flower. The goal is to train a computer to do what comes naturally to humans, which is to comprehend what is included in an image and provide insight from it. This project is interested in fine-tuning pre-trained machine learning models as well as training a flower recognition model from scratch. It’s essential to not only aim for novel and innovative methods but also conduct in-depth research on existing methods. This could lead to better insights and new discoveries. 121 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 2.0 LITERATURE REVIEW Lv et al. (2021) has proposed a flower classification model based on saliency detection and optimised VGG-16 deep neural network model tested on Oxford Flower-102 data set. To improve the model, optimization algorithm of stochastic gradient descent was done which can also reduce resource consumption and training time. Dropout method was used to reduce model overfit by randomly discarding training information. The use of transfer learning techniques was also done to solve the problem of insufficient image data which can also reduce model training time. This model shows an accuracy of 91.9%, which is higher than other traditional methods for image classification tasks and proves the feasibility of flower identification. Cibuk et al. (2019) employed pre-trained DCNN models for feature extraction. They chose two popular DCNN models, AlexNet and VGG-16, and concatenated features from both models to construct efficient feature sets. They used the minimum Redundancy Maximum Relevance (mRMR) model to act as a feature selection algorithm. Then, a support vector machine (SVM) classifier with Radial Bases Function (RBF) kernel is employed to classify the flower species using the extracted features. Their experimental results showed that they managed to achieve a 96.39% and 95.70% accuracy performance for the Flower-17 and Flower-102 dataset respectively. Gogul and Kumar (2017) has proposed a flower classification approach based on the Inception-v3 model of the Tensorflow platform, using transfer learning technology to retrain the flower categories. They have tested this approach using three models, the Inception-v3 model, the Xception model and the OverFeat model. On top of that, they have used a machine learning classifier such as Logistic Regression or Random Forest on top of the CNN models to increase the accuracy rate. Their approach minimizes the hardware requirements needed to perform the computationally intensive task of training a CNN. This approach outperforms all the handcrafted feature extraction methods such as Local Binary Pattern (LBP), Color Channel Statistics, Color Histograms, Haralick Texture, Hu Moments and Zernike Moments. This paper yields impressive Rank-1 accuracies of 73.05%, 93.41% and 90.60% using OverFeat, Inception-v3 and Xception architectures, respectively as Feature Extractors on Flower-102 dataset. Liu et al. (2016) proposed a flower classification approach using a convolutional neural network to extract features. They have also obtained the luminance map which is generated by converting RGB pixels to YUV, and the brightness of the color is extracted from the Y component, which allows better performance as flowers have high brightness. They use a regional contrast based salient object detection algorithm to compute a bottom-up saliency map, which simultaneously evaluates global contrast difference and spatial weighted coherence scores. The algorithm is simple, efficient, naturally multi-scale, and produces full- resolution, high-quality saliency maps which enhances the performance. They have achieved 76.54% accuracy in their dataset, and 84.02% in the Oxford Flower-102 dataset. 2.1 Existing Method SqueezeNet was selected as the model to be used for this flower recognition project. SqueezeNet is an innovative convolutional neural network which has 112 times fewer parameters than another CNN, Alexnet, while also maintaining an accuracy top-5 performance comparable to that of AlexNet. Being a small model, SqueezeNet is more amenable to on-chip implementations on FPGAs (Iandola et al., 2016). Research has been done using the SqueezeNet model in various use cases, and the results were promising. Sayed, 122 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Soliman and Hassanien (2021) implemented a model to predict melanoma skin cancer evaluated on ISIC 2020 and ISIC 2019, and uses a SqueezeNet model optimised with a bald eagle search (BES) optimization to find the best hyperparameter. The proposed melanoma skin cancer prediction model obtained an overall accuracy of 98.37%, specificity of 96.47%, sensitivity of 100%, f-score of 98.40%, and area under the curve of 99%. The experimental results showed the robustness and efficiency of the proposed model compared with VGG-19, GoogleNet, and ResNet50. ResNet, short for Residual Networks, allows engineers to train hundreds or even thousands of layers while still achieving impressive results. In 2015, this model won the ImageNet challenge. It has been studied that the increasing training error of deep neural networks is due to the network's initialization, optimization function, or one of the most well-known problems, the vanishing gradient problem (He et al., 2016). It is a problem that happens during the training of artificial neural networks using gradient-based learning and backpropagation. Gradients are known and used to update the weights in a network during backpropagation. However, sometimes the gradient becomes vanishingly small, effectively preventing the weights from changing values. This causes the network to stop training because the same values are propagated over and over again, resulting in no useful work being done. Residual neural networks are used to solve such problems. ResNet employs skip connections to add the output of an earlier layer to a later layer, thereby mitigating the vanishing gradient problem (He et al., 2016). VGG is an innovative object-recognition model which supports up to 19 layers. It is pre- trained with ImageNet datasets and is still able to outperform with other unseen datasets which make it one of the most used image recognition architectures. There are multiple variants for the VGGNet including VGG-16 and VGG-19, where these variants only differ in the total number of layers in the neural network. Multiple research has been done by using the VGG-19 model and impressive results were obtained. Victor Ikechukwu et al. (2021) has performed experiments using ResNet-50, ChexNet, VGG-19 and their own proposed Iyke- Net to identify pneumonia from chest x-ray images, where VGG-19 achieved a high accuracy of 93.5%, coming in close second after their proposed Iyke-Net which is 93.6% accurate. 3.0 METHODOLOGY This section talks about the datasets used and the design of the models. 3.1 Datasets The first dataset is the Kaggle flower recognition dataset consists of 4242 images from flickr, google images and yandex images (Mamaev, 2021). The images are divided into five classes: daisy, tulip, rose, sunflower and dandelion. There are about 800 images for each class, each image about 320x240 pixels. The photos are not reduced to a single size but come with different proportions. The Oxford Flower-102 dataset consisting of 102 more specific flower categories is also used (Nilsback & Zisserman, 2008). The images are flowers that are commonly found in the United Kingdom. There are between 40 to 258 images for each class, where each image has various scales, pose and light. There are large variations within the same category and several very similar categories which increase the difficulty of classification. The dataset has a total of 8189 images. 123 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 3.2 Model Architecture The first model is inspired by the original SqueezeNet model. In basic, it is a partial implementation of the original SqueezeNet (Iandola et al., 2016). It makes use of the Fire modules architecture as designed by the original authors. A fire module consists of a squeeze convolution layer of only 1x1 filters, which feeds into an expand layer that has a mix of 1x1 and 3x3 convolution filters. The liberal use of 1x1 filters greatly reduces the parameters as a 1x1 filter has 9 times fewer parameters than a 3x3 filter. The squeeze layer decreases the number of input channels to 3x3 filters, also greatly reducing the parameters in the layer. Thus, SqueezeNet is able to maintain reasonable accuracy while being more than 50 times smaller than another model, AlexNet and exceeding AlexNet’s top-1 and top-5 accuracy on the ImageNet dataset. This particular implementation is a stripped down version of SqueezeNet with fewer layers. It consists of one input into a conv2d layer, followed by batch normalization, the first fire module, the first MaxPooling2D layer, the second fire module, the second MaxPooling2D layer, the third fire module, the first GlobalAveragePooling2D layer and the final Dense layer with softmax activation to obtain the categories to be predicted. ResNet50 is chosen for transfer learning because of its lower computational power and promising accuracy compared to different ResNet variants like ResNet18, ResNet34, ResNet101 and ResNet152 (He et al., 2016). Firstly, the pre-trained ResNet50 model (resnet50 weights tf dim ordering tf kernels.h5) is downloaded from Github (Fchollet, 2016) and uses the weights from this downloaded model which trained from the imagenet datasets. Then, the first layer of ResNet50 is frozen and makes it non-trainable. If the first layer is trainable, the model may take a long time during the training process because it will have more trainable parameters. In this project, the pre-trained model weights should not be retrained because they are the advantage while taking the transfer learnt model. Besides, when a pre-trained network is used for transfer learning, additional dense layers shall be added at the end of the pre-trained network in order to learn which combination of the previously learned features helps in recognising the objects in the new dataset. Thus, additional layers were implemented after the output of the ResNet50 model. A total number of 7 layers were added in this project, that is the flatten layer, batch normalization layers, two customised layers with size 2048 and 1024, accompanied with “Rectified linear unit” (ReLU) activation function and the softmax layer. The flatten layer converts data into a 1-dimensional array for input to the next layer, and batch normalisation is a layer that allows each layer of the network to learn more independently. Finally, a softmax layer is included as the output layer in order to predict fixed types of flowers whereby only 5 or 102 classes of classification are produced, depending on the dataset. The 5 classes are produced to predict the 5 types of flowers for the 5 category dataset, which are daisy, tulip, rose, sunflower and dandelion. So, after applying this layer, a transfer learning model is created that can classify the input images into various types of flowers based on predictions from the pre-trained ResNet50 model. VGG-19 is composed of 16 convolutional layers, 5 pooling layers, 3 fully-connected layers and a final layer of softmax function (Simonyan & Zisserman, 2015). The matrix was shaped (224, 224, 3) as the fixed input size of (224x224) RGB image is passed into this network. Kernel size of (3x3) instead of large kernels with stride of 1 pixel is used to cover every part of the image, whereas a 2x2 pixel window with a stride of 2 pixels is used to perform max pooling. The multiple layers of small kernels are able to effectively cover the 124 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 images without the use of large kernels such as 11x11 kernel in AlexNet and 7x7 kernel in ZFNet. Therefore, the number of parameters is also reduced and the overfitting problem is reduced. All hidden layers are equipped with ReLU to introduce non-linearity for better classification compared to previous models which used tanh or sigmoid functions. In this project, the pretrained weights are used by setting the parameter weights to imagenet. The default classifier is removed by setting include_top to false so a new classifier can be created. The first 19 layers are frozen to prevent the weights from being modified. Similar to the ResNet50 model, additional layers are added after the pretrained model. A max-pool layer is added to down-sample the input features. A flattening layer is then added before a dense layer with a softmax function since the dense layer accepts 2D input. 3.3 Logical Flow Figure 1. Logical Flowchart. Data Loading: The flowers dataset consists of images of flowers with different class labels and stored in respective directories. This stage is to load the data from their directories and concatenate it into one dataframe. As a result, the original flower images end up with an image dimension of 244x244, which minimises image dimensions while maintaining image readability with efficient computational complexity and accommodating the input shape for the pre-trained models. For the 5 category dataset, there are 4242 images and 5 class labels. For the 102 category dataset, there are 8189 images and 102 class labels. 125 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Data Understanding: The flowers dataset contains examples of labelled flower images. Each example includes a JPEG flower image as well as the class label. The exploration in the image data helps to validate the class distribution of each type of flower and ensures a balanced dataset, preventing an imbalanced dataset that leads to poor prediction. Besides, data visualization is required to sample and examine the input data to ensure image readability by randomly previewing the 10 images in the dataframe via 2D representation with the new image dimensions. Data Labelling: This stage is used to transform categorical data (textual data) into numerical values for the prediction functions so that the deep learning predictive models can understand. Label encoding technique is performed to convert categorical values to numbers in this step. Model Creation: The same optimizer and loss function is used to compile all models. Cross-entropy is the loss function chosen to evaluate a set of weights in multi-class classification problems for flower recognition. Furthermore, the Adam optimizer with the default learning rate (0.01) is used to search through different weights for the network. Finally, because this is a classification problem, classification accuracy is collected and reported, which will be defined via the metrics argument. Data Shuffling: This stage is used to redistribute the training and testing data samples in the dataset to ensure that each data sample produces an \"independent\" change on the model, without being influenced by the points that came before it. Since the images were added sequentially from subfolders into the dataframe during data loading, data shuffling is required. Otherwise, the model can only learn what is \"daisy\" from the first 800 images, which does not optimise the model's parameters. The seed is set to 100 and is applied to both images and labels to ensure that each image matches the correct label. Data Augmentation: This stage is used to increase the amount of data by adding slightly modified copies of already existing data or newly created synthetic data from existing data. This refers to randomly changing the images in ways that shouldn’t impact their interpretation, such as horizontal flipping, zooming, and rotating. Through data augmentation, it acts as a regularizer and prevents overfitting when training a machine learning model. It is the technique used to overcome the problem of overfitting by creating more data and making the model generalize well on the unseen data. Model Training: The model is using the predefined train-test split for the Oxford Flower- 102 dataset, with 1020 training samples and 6149 testing samples. The 5 category dataset is split into train-test sets using a ratio of 80:20. Different batch sizes and epochs are adjusted in different models in order to achieve the optimal result. This stage generates tensor image data in batches, which will be looped over for both training and testing. The neural network in each model takes in inputs, which are then processed in hidden layers using weights that are adjusted during training. Then the model spits out a prediction. The weights are adjusted to find patterns in order to make better predictions. Model Evaluation: In this stage, a learning curve and confusion matrix is presented to evaluate the model performance. There are four aspects to be compared: training accuracy, training loss, validation accuracy and validation loss in every epoch. Usually, with every epoch increasing, the loss should be going lower and accuracy should be going higher. The validation loss and validation accuracy measures were calculated after the model had gone through all the data. So the network had been fully trained when these scores were calculated. On the other hand, the confusion matrix is a summary of classification problem prediction 126 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 results. The number of correct and incorrect predictions for each flower class is summarised with count values and broken down by each class. 4.0 RESULTS AND DISCUSSION This section shows the results obtained for each model, on both datasets. The metric used is the validation accuracy as metrics like training accuracy do not reflect real-life performance. The validation accuracy is calculated by taking the True Prediction/Total Number of Predictions using the validation dataset. The models were validated with N = 1020 samples for the Oxford Flower-102 dataset and N= 848 samples for the 5 category dataset. Table 1. Validation accuracy for each model for the two datasets. Validation Validation Status Accuracy (5 Accuracy Category) Rejected (102 Rejected SqueezeNet Inspired Model 77% Category) Accepted ResNet50 67% VGG19 88% 67% 42% 84% Based on our results, the partially pre-trained VGG-19 model performed the best on both datasets, achieving a high validation accuracy of 88% and 84% on the 5 Categories and Flower-102 dataset respectively. This may be the result of the simple structures of VGG-19 and its hidden layers with ReLU function that can better introduce non-linearity for better classification compared to other models. The smaller number of features produced also contributes to the better generalization of the model. The SqueezeNet Inspired model had the second highest validation accuracy of 77% and 67% on the 5 Categories and Flower-102 dataset respectively. This might be due to the fact that while this SqueezeNet Inspired model is far less complex and has less performance than both other models, this model is not pretrained. Instead, the whole model was trained solely on the two datasets individually for each scenario, and therefore was able to perform better than the partially pre-trained ResNet50. This is a lightweight model, even when compared to the original SqueezeNet implementation, and therefore it should be expected that there were sacrifices in terms of performance. Finally, the ResNet50 model has the least performance, achieving 67% and 42% on the 5 Categories and Flower-102 dataset respectively. This could be due to the ResNet50 model's architecture being overly complicated for this task, leading to poor generalisation. The selection of optimal learning rate, batch size and identifying the best freezing layer is playing an important role for the better performance of the model. 5.0 CONCLUSION The objectives are fulfilled as a model using VGG19 has been created that can perform flower classification tasks with 89% accuracy with 5 categories of flowers, and 84% accuracy on 102 species of flowers. However, there are limitations in this research such as limitations in hardware. As many of the algorithms and models in this project are all computationally intensive. This project is also more concerned with which models perform better on the two datasets of flowers, and not with exploring why the models perform the way that they do. 127 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Another potential avenue for exploration is the other pretrained models like AlexNet, VGG16 and various other models, which are not used in this project. More comprehensive research comparing even more models can be done in the future to ensure completeness. 6.0 ACKNOWLEDGEMENTS Authors thank the Faculty of Computing and Information Technology, Tunku Abdul Rahman University College for financial support and resources to carry out this project. REFERENCES Cıbuk, M., Budak, U., Guo, Y., Cevdet Ince, M. and Sengur, A., (2019) ‘Efficient deep features selections and classification for flower species recognition’. Measurement, 137, pp.7-13. Fchollet, (2016). Release VGG16, VGG19, and ResNet50 · fchollet/deep-learning-models., GitHub. Available at: https://github.com/fchollet/deep-learning- models/releases/tag/v0.1 (Accessed: 10 September 2021) Gogul, I. and Kumar, V. (2017) ‘Flower species recognition system using convolution neural networks and transfer learning’. 2017 Fourth International Conference on Signal Processing, Communication and Networking (ICSCN). He, KM., Zhang, XY., Ren, SQ., Sun, Jian., (2016) Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p.770-778. Iandola, F. , Han, S., Moskewicz, M., Ashraf, K., Dally, W. and Keutzer, K. (2016) ‘SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size’. Liu, Y., Tang, F., Zhou, D., Meng, Y. and Dong, W. (2016) ‘Flower classification via convolutional neural network’. 2016 IEEE International Conference on Functional- Structural Plant Growth Modeling, Simulation, Visualization and Applications (FSPMA). Lv, R., Li, Z., Zuo, J., Liu, J. (2021) ‘Flower Classification and Recognition Based on Significance Test and Transfer Learning’. 2021 IEEE International Conference on Consumer Electronics and Computer Engineering, ICCECE 2021. Institute of Electrical and Electronics Engineers Inc., pp. 649–652. Mamaev, A. (2021). Flower Recognition, Kaggle [Online]. Available at: https://www.kaggle.com/alxmamaev/flowers-recognition (Accessed: 19 April 2021) Nilsback, M., Zisserman, A. (2008). 102 Category Flower Dataset, Visual Geometry Group [Online]. Available at: https://www.robots.ox.ac.uk/~vgg/data/flowers/102/ (Accessed: 19 April 2021) Simonyan, K., Zisserman, A. (2015) ‘Very deep convolutional networks for large-scale image recognition’, 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, International Conference on Learning Representations, ICLR. 128 of 225 ICDXA/2021/12 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 THE SCIENCE OF EMOTION: MALAYSIAN AIRLINES SENTIMENT ANALYSIS USING BERT APPROACH Huay Wen Kang1*, Kah Kien Chye1, Zi Yuan Ong1 and Chi Wee Tan1 1 Faculty of Computing and Information Technology, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia *Corresponding author: [email protected] ABSTRACT Sentiment analysis has grown to be one of the most active research areas in Natural Language Processing (NLP). Sentiment analysis, also known as opinion mining, uses a series of methods, techniques and tools to study people’s opinions, views and sentiment towards a wide range of topics such as products, services, events and issues. In the airline industry, millions of people today use social networking sites such Twitter, Skytrax, TripAdvisor to express their emotions, opinions, and share information about the aircraft service. It is a hidden gem to the airline company to gain valuable insight from this data and have the broadest possible view into what people are saying about the airline’ brand online. Hence, this paper explores six different sentiment analysis models: Random Forest, Multinomial Naive Bayes, Linear Support Vector Classifier, Ensemble Method, Bidirectional Long Term Short Memory (Bi-LSTM) and BERT model, in order to determine and develop the best model to be used. The best model was then used to determine the social status, company reputation, and brand image of Malaysian airline companies. In conclusion, the BERT model was found to perform the best out of the six models tested, scoring an accuracy of 86%. Keywords: Supervised Learning, Ensemble Learning, Deep Learning, Transfer Learning, Airline Sentiment 1.0 INTRODUCTION Social media generates a large amount of sentiment-rich data in the form of tweets, status updates, comments, reviews and others. The age of the Internet has changed the way people express their thoughts and opinions. It is now done primarily through blog posts, online forums, product review websites, social media and so on. For large airlines with thousands of daily references on social media, news sites and blogs, it’s extremely difficult to manage and track these references manually. That’s where natural language processing comes in. Therefore, the objective for this project is to conduct research and analysis on the social status, company reputation, and brand image of Malaysian airline companies, specifically Malaysia Airlines, AirAsia and Malindo. This project investigates various techniques and algorithms in terms of supervised learning, ensemble learning, deep learning, and transfer learning in order to develop an airline sentiment analysis model. Besides, different n-gram features are evaluated for their performance to determine which one performs the best. 129 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 2.0 LITERATURE REVIEW Sousa et al. (2019) has evaluated BERT in the task of stock market analysis to predict the following movements of the Dow Jones Industrial (DJI) Index. The data consists of 582 financial news that was crawled from various news websites such as CNBC, Forbes and New York Times using Selenium tool and the dataset was manually annotated as positive, negative and neutral sentiment. Tokenization was done using WordPiece to transform each document into a token sequence. The parameter for attention head was set to BERT BASE as it is smaller and is able to cater to their limited computational power. This is followed by a 10- fold cross validation before training the models using labeled data. The results were then compared to Support Vector Machines, Naive Bayes and Convolutional Neural Network, where BERT clearly outperformed other methods with the highest F1-score of 72.5%. This analysis was used as an indicator of falling and rising of the economy in the day. Ortigosa, Martín and Carro (2014) implemented a new method for sentiment analysis in Facebook, which extracts information from the messages about sentiment polarity, and then models the users’ usual sentiment polarity and to detect significant emotional changes. This method was then implemented in SentBuk, a Facebook application that retrieves user messages and classifies them according to polarity and shows the results to the users through an interactive interface, while also supporting emotional change detection, friend’s emotion finding, user classification, and statistics. The classification method implemented in SentBuk follows a hybrid approach: it combines lexical-based and machine-learning techniques. The results obtained through this approach show that it is feasible to perform sentiment analysis in Facebook with high accuracy (83.27%). Andrew et al. (2011) proposed a model for sentiment analysis using a mix of unsupervised and supervised techniques to learn word vectors capturing semantic term- document information as well as rich sentiment content. The model’s probabilistic foundation gives a theoretically justified technique for word vector induction as an alternative to the overwhelming number of matrix factorization-based techniques commonly used. The model is parametrized as a log-bilinear model following successes in using similar techniques for language models. The topical component of the model was parameterized in a manner that aims to capture word representations instead of latent topics. Their experiments showed their method performing better than Latent Dirichlet Allocation (LDA) which models latent topics directly. Their unsupervised model incorporates sentiment information by leveraging the abundance of sentiment-labeled texts online to yield word representations that capture both sentiment and semantic relations. Two tasks of sentiment classification were carried out using existing datasets as their own larger dataset. The model is shown to be highly flexible and can be used to characterize a wide variety of annotations, and thus is broadly applicable in the growing areas of sentiment analysis and retrieval. 2.1 Existing Method Naive Bayes are a family of classifiers based on Bayes’ popular probability theorem, and are well suited for creating simple but powerful models, particularly in the area of textual classification. Naive Bayes is easy and quick to implement, and is less faulty compared to more complex and slower algorithms. Naive Bayes models require comparatively little data to train, and estimate the required parameters (Abbas et al., 2019). There are several types of Naive Bayes Classifier, which are Multinomial Naive Bayes (MNB), Bernoulli Naive Bayes (BNB) and Gaussian Naive Bayes (GNB). A Naive Bayes classifier is essentially a probabilistic machine learning model used for classification, using the Bayes theorem, which 130 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 describes the probability of an event based on the prior knowledge of the conditions that might be related to the event. Multinomial Naive Bayes was selected instead of the other Naive Bayes types because MNB is an improved version of Multivariate Bernoulli Naive Bayes model (BNB) and takes into account word frequency and information and thus obtains better accuracy. Support vector machine (SVM) is a learning technique that excels at sentiment analysis, as this algorithm can significantly reduce the need for labelled training instances in both the standard inductive and transductive settings (Phienthrakul, 2009). SVM draws that hyperplane by transforming data with the help of mathematical functions called “Kernels”. Types of Kernels are linear, sigmoid, RBF, non-linear, polynomial, etc. The performance of SVM is determined by the kernel function used. As a result, if the appropriate kernel is chosen, the efficiency of classification will be improved. In the past, research on sentiment classification with SVM and multiple kernel functions was conducted, and promising results with high accuracy were obtained. However, SVM has evolved into various versions in order to adapt to new changes and be more flexible in today's world. Various extensions of SVM such as support vector classifier, multiclass SVM, transductive support-vector machines, structured SVM, bayesian SVM, and others had been proposed by the researchers over the time. Hence, other combination methods may improve the efficiency of sentiment classification compared to traditional SVM. This led to the idea of using a different implementation of SVM, that is the support vector classifier in this airline sentiment project. The Random Forest algorithm consists of a large number of individual decision trees that operate in an ensemble manner, resulting in a forest of trees. Rather than using best split among all variables to split each node, Random Forest randomly selects a subset of predictors and chooses the best among them. The tree grows with randomly selected features and is not pruned, which results in Random Forest having amazing accuracy (Breiman and Cutler, 2004). For Random Forest to perform well, each individual tree needs to have low correlation with each other. Random Forest has proved to be a suitable classifier to perform sentiment analysis in multiple research projects. Based on a sentiment analysis on Malaysian mobile digital payment applications research done by Balakrishnan et al. (2020), Random Forest achieved the highest accuracy of 75.62% and f1-score of 71.99% amongst other algorithms including Support Vector Machine, Naive Bayes and Decision Tree. Likewise, Hedge and Padma (2017) also obtained an accuracy of 72% when using Random Forest to perform sentiment analysis for mobile product reviews in Kannada. The ensemble method combines the results of several models to produce better predictive performance compared to a single model. This is able to eliminate the possibility of overfitting while boosting the overall performance (Yu et al., 2010). Besides that, ensemble methods are also commonly used to solve the class imbalance problem using multi- objective optimization algorithms (Yang et al., 2020) and also minimize the number of features better than existing ensemble algorithms such as AdaBoost and Gradient Boosting. In Wan and Gao’s (2016) experiment of using an ensemble sentiment classification system for airline services analysis, the ensemble classifier using a majority vote method with five classifiers has obtained the highest accuracy of 84.2% in comparison to other single classifiers. The recurrent neural network (RNN) family of models, especially the LSTM networks, has been shown to be the most effective sequence models used in practical applications (Goodfellow, Bengio and Courville, 2016). Bi-directional LSTMs are used to learn from both 131 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 the forward and backward time dependencies. In a Bi-derectional LSTM, each unit is split into two units having the same input and connected to the same output. One unit is used for the forward time sequence and the other for the backward sequence. Thus, Bi-directional LSTM is useful when learning from long spanning time-series data, and shows improved results without increasing the training time. BERT (Bidirectional Encoder Representations from Transformers) model was introduced in 2018 to quickly and effectively create a high-quality model with minimal effort and training time using the PyTorch interface, regardless of the specific NLP task, and produce state of the art results. Recently, a sentiment analysis on the impact of coronavirus on social life using the BERT model achieved 94% validation accuracy on the collected data sets (Singh et al., 2021). In short, BERT is one of the most powerful NLP models available at the moment, requiring only a small amount of data while achieving state-of-the-art results with minimal task-specific adjustments for a wide range of NLP tasks such as named entity recognition, language inference, semantic similarity, question answering, and classification like sentiment analysis. 3.0 METHODOLOGY This section talks about the datasets used and the design of the models. 3.1 Datasets The dataset is the data from Twitter about sentiment towards US major airlines (US Airways, United, Virgin America, Delta, Southwest, American) obtained from Kaggle. There are 9178 negative tweets, 3099 neutral tweets and 2363 positive tweets respectively. It is split into a train and test set with an 80:20 ratio for model training. The models trained using this dataset are expected to be generalized and able to handle all airline reviews regardless of company or country as long as they are all in English. The dataset for three Malaysia airlines (Malaysia Airlines, AirAsia, Malindo) was scraped from TripAdvisor using UiPath from July 2021 for model deployment purposes. There are around 460 pieces of data for each class. The data scraped includes the text reviews that are in English and the username of the person who posted the review. 3.2 Text Preprocessing For traditional machine learning methods, data preprocessing was done on the training data before being used to train the models. Firstly, emoticons and emojis are not removed and are converted to their textual meaning using the emot library (Shah & Rohilla, 2018) as emoticons and emojis carry sentiment. Next, all words in the message are converted into lowercase in order to normalize the text. To reduce noise, symbols, punctuations, hyperlink, extra whitespace, new lines along with digits are removed. Furthermore, stopwords are removed and lemmatization is done to return the words back to their root word since the word would be disrupted by an irrelevant inflection like a simple plural or present tense inflection. On the other hand, BERT is trained on raw text with no text preprocessing such as stopword removal, lemmatization and others. This is because the model has a fixed vocabulary and the BERT tokenizer has a specific way of dealing with out-of-vocabulary words. As a result, the previously preprocessed data is omitted, and instead the bert-tokenizer is used for text formatting. Feature selection is done to transform the text into unigram word vectors using CountVectorizer. Next, due to the data imbalance between positive, neutral and negative 132 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 sentiment, resampling techniques are used to address the issue. This is because class imbalance might affect the accuracy and performance of the models. Therefore, random undersampling is done to reduce the majority class randomly down to the desired ratio against the minority class, and combined with using SMOTE to do oversampling, which is oversampling by creating \"synthetic\" examples of the minority class. The undersampling and SMOTE both use the default sampling strategy, which for undersampling is “not minority” and resample all classes but the minority, and for SMOTE is “not majority”, which oversamples all classes but the majority. Both will equalise the unequal classes. After the process, the training dataset will have N = 21867 samples from the initial 11712 samples. 3.3 Model Applications Naive Bayes (NB) is based on Bayes theorem that can be used for classification challenges. Multinomial Naive Bayes (MNB) is a variant of NB which is a probabilistic learning method mostly used in Natural Language Processing (NLP). MNB performs calculation of probabilities based on probabilities of causal factors and is useful to model feature vectors where each value represents the number of occurrence of a term, where in the case of NLP a text can be considered as a particular instance of a dictionary and the relative frequency of all terms provide enough information to infer a belonging class. The hyperparameters for this algorithm are the additive laplace smoothing parameter, which is intended to solve the problem of the zero probability in the Naive Bayes algorithm, whether to learn class prior probabilities or not and prior probabilities of the classes. For the additive laplace smoothing parameter, using higher alpha values will push the likelihood towards a value of 0.5. Therefore, it is preferable to use lower alpha values. The line below shows the optimized parameters for the model after hyperparameter tuning: MultinomialNB(alpha=0.5) A simple, linear support vector classifier (SVC) is proposed to classify the airline data into three different classes. SVC is a different implementation of the same algorithm, that is the SVM which is implemented in terms of liblinear rather than libsvm (Asif et al., 2020). It is just a thin wrapper around libsvm but has more flexibility in the choice of penalties and loss functions and should scale better to large numbers of samples. It supports both dense and sparse inputs and also handles multi classes. The classifier will take into account each unique word present in the sentence, as well as all consecutive words which are suitable for text classification. By deploying this mechanism, it segregates the airline data into three categories by using a linear hyperplane to classify the posts and comments into positive, negative and neutral classes. With that, the text involved in the airline dataset is identified through sentiment analysis. Common features such as gamma are scale at 0.5, regularisation parameter (c) of 100, and the use of probability estimates in conjunction with a balanced mode after parameter tuning. The call of the library from the Sklearn module is made as follows: SVC(C=100, class_weight='balanced', gamma=0.5, kernel='linear', probability=True) The hyperparameters random forest include the number of decision trees in the forest, number of features considered by each tree when splitting a node, number of levels in each decision tree, min number of data points allowed in a leaf nod, min number of data points placed in a node before the node is split and number of trees in the forest. The parameters of a random forest are the variables and thresholds used to split each node learned during training. The line below shows the optimized parameters for the model after hyperparameter tuning: RandomForestClassifier(max_depth=100, max_features='sqrt', min_samples_leaf=2, min_samples_split=10, n_estimators=800) 133 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 For the voting classifier, the previous three algorithms (Multinomial Naive Bayes, Random Forest and LinearSVC) with their respective optimized parameters set are passed into the classifier and used soft voting, which predicted the class with the largest summed probability from models. Bidirectional LSTMs are an extension of traditional LSTMs, which train two LSTM instead of one on the input sequence, with the second LSTM trained on the reversed copy of the input sequence, which provides additional context to the network and results in faster and even fuller learning on the problem. Before feeding input into the Bidirectional LSTM, the data was preprocessed. In summary, the sentiments were converted into a binary class matrix, tokenizing the text, converting the text into an encoded form, and finally, padding all the encoded text to the same length. The proposed Bidirectional LSTM model is a simple one, consisting of only 7 layers, with one embedding input layer and one embedding layer, followed by one Convolutional layer, one max pooling layer, one bidirectional LSTM layer, one dropout layer and finally one dense layer to get 3 classes for classification. BertForSequenceClassification model with an added single linear layer on top for classification that serves for sentence classifier purposes is used in this project. After the airline data sets are fed in, the entire pre-trained BERT model and the additional untrained classification layer is trained on the specific task, that is the multi-class classification. For fine-tuning, most model hyperparameters remain in the default value, with the exception of the batch size, learning rate, and the number of training epochs. The author discovered a range of feasible values that function well across all NLP tasks, hence the batch size is set to 32, followed by a 2e-5 learning rate (Adam) from the recommended range values and a maximum epoch value of 4. Besides, the model will be trained under the PyTorch framework. The model will undergo the typical Pytorch training cycle and loop for 4 times, which includes iterating through the mini-batches, performing a feedforward pass for each batch, computing the loss, performing backpropagation for each batch, and finally updating the gradients. In short, this fine-tuned model will leverage the power of the pre-trained transformers (BERT) model and PyTorch framework. 4.0 RESULTS AND DISCUSSION The following table shows the results of all models. Unigram features were used for Multinomial Naive Bayes, Linear Support Vector Classifier, Random Forest and Voting Classifier. The testing dataset has N = 2928 samples. Table 1. Results for all models Status Accuracy Precision Recall F1-Score Multinomial Naive Bayes 70.4% 77.7% 78.2% 77.9% Rejected Linear Support Vector Classifier 66.7% 74.0% 73.8% 73.9% Rejected Random Forest 66.5% 76.3% 77.3% 76.5% Rejected Voting Classifier 68.4% 79.2% 80.2% 78.7% Rejected Bidirectional LSTM 77.0% 77.0% 77.0% 77.0% Rejected BERT 86.0% 86.0% 86.0% 86.0% Accepted 134 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Among the traditional machine learning techniques (Multinomial Naive Bayes, Linear Support Vector Classifier and Random Forest), Multinomial Naive Bayes achieve the highest results. Hence it is further used to test using bigram and trigram features. Table 2. Results for Multinomial Naive Bayes using different n-gram orders Accuracy Precision Recall F1-Score Unigram 70.4% 77.7% 78.2% 77.9% Bigram 53.4% 64.9% 66.9% 65.6% Trigram 41.4% 61.3% 65.8% 59.1% Based on the experimental result, the BERT model provided the best results among the six models with an accuracy of 86.0%. It shows that transfer learning approaches outperform traditional machine learning algorithms in airline sentiment analysis. Experiments show that BERT is superior to supervised text classification without human supervision, such as the standard preprocessing process, which includes decapitalization, punctuation removal, stopword removal, and emoji conversion to text. The Bidirectional LSTM performed the second best, with an accuracy of 77.0%. This result is better than all the traditional machine learning methods tried and therefore proves the hypothesis that deep learning models perform better than traditional machine learning algorithms. This may be due to the fact that this Bidirectional LSTM also takes into account backward and forward time-dependencies. Among the different n-gram models tested for Multinomial Naive Bayes, unigram features lead to the highest accuracy (70.4%). The accuracy further decreases with every increase in n-gram. This may be due to when the n-gram length increases, the frequency of any given n- grams will decrease and it may not generalize well to a different data set, hence resulting in lower accuracy. 5.0 CONCLUSION In conclusion, this paper has explored various sentiment analysis models, and have found that the best machine learning model out of the four selected is Multinomial Naive Bayes with an accuracy of 70.4%. It was also shown that the Unigram model of the Multinomial Naive Bayes performs the best when compared with Bigram and Trigram models. It is shown by the results that Deep Learning methods perform better than traditional machine learning methods, with both the Bidirectional LSTM (77% accuracy) and BERT model (86% accuracy) outperforming all traditional machine learning methods. Finally, we have determined that out of all six models tested, the BERT model performs the best. A high accuracy model for sentiment analysis has successfully developed and applied to a new airline reviews dataset. Based on the crawled data, out of 1371 reviews, the majority were categorized as negative (N=1118, 81.55%), 226 positive (16.48%) and 27 neutral (1.97%). This suggests that most Malaysia’s Airline’s customers are not satisfied with the airline services. Companies can have better directions and make better business decisions through the insights obtained from sentiment analysis, such as providing better training for their flight crews if negative sentiments towards their services are found. However, the crawled data used in this project is relatively small, such a small and sparse dataset might not be truly representative of the wider population, and also restricts 135 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 several types of data analysis such as trend analysis due to lack of data. As such, this project can be improved by crawling more data to be used for the dataset and reducing the sparsity of the data, so that it is more representative of the actual population and more complex analysis can be carried out. 6.0 ACKNOWLEDGEMENTS Authors thank the Faculty of Computing and Information Technology, Tunku Abdul Rahman University College for financial support and resources to carry out this project. REFERENCES Abbas, M., Memon, K., Jamali, A., Memon, S. and Ahmed, A. (2019) ‘Multinomial Naive Bayes Classification Model for Sentiment Analysis’. International Journal of Computer Science and Network Security, 19(3), pp.62-67. Andrew, L., Raymond, E., Peter, T., Dan, H., Andrew, Y., Christopher, P. (2011) ‘Learning Word Vectors for Sentiment Analysis’. Stanford University. Asif, M., Ishtiaq, A., Ahmad, H., Aljuaid, H. and Shah, J. (2020) ‘Sentiment analysis of extremism in social media from textual information’. Telematics and Informatics, 48, p.101345. Balakrishnan, V., Selvanayagam, P.K., Yin, L.P. (2020) ‘Sentiment and Emotion Analyses for Malaysian Mobile Digital Payment Applications’. ACM International Conference Proceeding Series. Association for Computing Machinery, pp. 67–71. Breiman, L., Cutler, A. (2004). ‘RFtools for Predicting and Understanding Data’. Interface Workshop 1–62. Goodfellow, I., Bengio, Y. and Courville, A. (2016) ‘Deep Learning’. MIT Press. Hegde, Yashaswini & Padma, S.K. (2017). Sentiment Analysis Using Random Forest Ensemble for Mobile Product Reviews in Kannada. 777-782. 10.1109/IACC.2017.0160. Neel Shah & Shubham Rohilla (2018). Description of the emot library, v2.2. Github [Online]. Available at: https://github.com/NeelShah18/emot (Accessed: 19 April 2021) Ortigosa, A., Martín, J. and Carro, R. (2014) ‘Sentiment analysis in Facebook and its application to e-learning’. Computers in Human Behavior, 31, pp.527-541. Phienthrakul, T., Kijsirikul, B., Takamura, H. and Okumura, M. (2009) ‘Sentiment Classification with Support Vector Machines and Multiple Kernel Functions’. Neural Information Processing, pp.583–592. Singh, M., Jakhar, A.K. and Pandey, S. (2021) ‘Sentiment analysis on the impact of coronavirus in social life using the BERT model’. Social Network Analysis and Mining, 11(1). Sousa, M.G., Sakiyama, K., Rodrigues, L.S., Moraes, P.H., Fernandes, E., Matsubara, E.T. (2019) ‘BERT for stock market sentiment analysis’, International Conference on Tools with Artificial Intelligence, ICTAI. IEEE Computer Society, pp.1597–1601. Wan, Y., Gao, Q. (2016). ‘An Ensemble Sentiment Classification System of Twitter Data for Airline Services Analysis’. 15th IEEE International Conference on Data Mining Workshop, ICDMW 2015. Institute of Electrical and Electronics Engineers Inc., pp. 1318–1325. Yang, K., Yu, Z., Wen, X., Cao, W., Chen, C, Wong, H., You, J. (2020). ‘Hybrid Classifier Ensemble for Imbalanced Data’. IEEE Transactions on Neural Networks and Learning Systems 31, 1387–1400. Yu, H., Lo, H., Hsieh, H., Lou, J., McKenzie, T., Chou, J., Chung, P., et al. (2010). ‘Feature engineering and classifier ensemble for KDD cup 2010’. JMLR: Workshop and Conference Proceedings 1–12. 136 of 225 ICDXA/2021/13 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 DIGITAL CULTURE: ONLINE SHOPPING ADOPTION AMONG COLLEGE STUDENTS IN MALAYSIA Chiun Wei Puah1, Weng Lam Eng1, Chun Hoong Tan1, Shuen Chen Tan1 and Tin Tin Ting2* 1 Faculty of Computing and Information Technology, Tunku Abdul Rahman University College, Kampus Utama, Jalan Genting Kelang, 53300, Wilayah Persekutuan Kuala Lumpur, Malaysia 2 Faculty of Information Technology, INTI International University, Negeri Sembilan, Malaysia *Corresponding author: [email protected] ABSTRACT Online shopping is getting popular in Malaysia during Covid-19 pandemic and country lock down since 2020. Researches are important in inspecting this arising digital culture in Malaysia in order to encourage the participation of customers in online shopping and reduce their cognitive deficiencies in the e-commerce, thus promoting economy growth. The purpose of this study is to examine the factors of the college students adopting this digital culture – online shopping behaviour. The relationship between online advertisement, product risk, delivery risk and information security in online shopping behaviour among college students are examined. Online questionnaire is designed in Google Form and data collected is analysed using PSPP utilizing one-way ANOVA and Cohen’s f effect size. The study revealed that online advertisement, product risk, delivery risk, and information security significantly affect online shopping behaviour of college students with large effect size. Keywords: Online shopping behaviour, Customer experience, Electronic commerce, Perceived risks, Online advertisement 1.0 INTRODUCTION Globalization, Internet access, Covid-19 pandemic, and country lock down are several main reasons that have coaxes digital culture – online shopping in Malaysia recently. The country has recorded a boost in online shoppers to 16.53 million (50% of the population) in August 2019 (ITA, 2019). The statistic continues to increase in 2020 following Covid-19 pandemic and country lock down starting in 18 March, 2020. Another main reason of the arising online shoppers is the increasing spending power of customers (Anurag and Jitesh 2019). Ease of access and ease of use have also contributed to the statistics. At the online shopping website, online customers may choose from over 30 million products in over 70 categories product. These e-retailers offer an electronic catalogue of items from which customers may choose things and compare them to other products before making a purchase. Online shopping offers several advantages, including worldwide reach, a wide range of items, and the necessary information. It saves time while purchasing products since it removes the time spent driving to and from a physical store. Customers may acquire items at any time, at the lowest possible price, and receive offers and discounts while purchasing products online. According to previous researches, factors such as advertisement, discounts, products and delivery quality provided in the online shopping websites significantly affect online shopping behavior (ITA, 2019; Lee et al., 2018; Onewo et al., 2020; Yeniçeri & Akin 2013; Ganapathi and Abu-Shanab, 2020; Zamzuri et al., 2018). It is an interest in this study to examine factors that might affect younger generation in Malaysia (college students) although they are not the main customers in the past due to spending power as compared to 137 of 225 ICDXA/2021/14 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 working adults. However, the number of young customers is increasing dramatically as technology advancing (Statista, 2020). Although online shopping brings many benefits, it also has disadvantages. For example, customer sensitive information might be exposed during the checkout session hence causing some potential danger. Customers cannot touch or feel the product they are willing to buy which makes the customers worry about paying for the wrong product and thus prefer tangible products (Mac, 2019). Meanwhile, customers might have wasted their time to search for those unrelated products or promotions that are misled by the advertisement (Kim et al., 2008). There are also some bad experiences such as delivery not available in certain regions or poor delivery service. Many new online shoppers are also worried about the security of payment information or personal private information. All these negative factors will affect customers' purchase behavior and also their user experience. The theory of planned behavior (TPB) is widely used in predicting purchase intention nowadays even among online shoppers. A person’s perceived behavior control is defined as one’s perception on the ease of carrying a specific behavior (Neeraj and Veena, 2016). For example, perceived ease of use is a proven factor that could predict purchase intention (Ahmad, 2017). However, the negative aspects are equally important in affecting one’s decision. This study is focusing in the negative aspects of the availability of resources, mainly in the perceived risk. Based on Emad (2013) research that focus in the negative aspects of perceived risk on online shopping behavior, the four (out of six) factors that are affecting online shopping behavior are: financial risk, product risk, delivery risk and information security. However, financial risk might not be applicable to college student as financial risk is referring to online credit card usage. On the positive aspect, according to Nur et al. (2020), online advertisement factor has the highest correlation with online shopping behavior compared to accessibility and entertainment factors. Therefore, this study will focus in four factors which are Online Advertisement, Product Risk, Delivery Risk, and Information Security in order to examine the new digital culture adoption among college students. Although there are many researches done on the factors affecting online shopping behavior, there is a limitation in a study that focus in college students. According to a survey on 5623 respondents, more than half of the respondents from age group 16 to 24 (56%) purchased more (Statista, 2020). Therefore, this group of customers are potential players in e- commerce. This study will analyze the possible factors affect online shopping behavior among college students (age 16 to 24). In addition, effect size of the factors is answered in the analysis. 2.0 LITERATURE REVIEW Online Shopping Behaviour refers to an individual's overall view and assessment of a product or service while buying online, which can be negative or positive (Shahzad et al., 2015). Meanwhile, Online Advertising is a selling strategy that involves the use of the Internet as a medium to generate website traffic and present marketing messages to the right customers (Khandare and Suryawanshi, 2016). Online advertising helps the organization or company to promote their products through brand recognition (Haider and Shakib, 2017). Vincent et al. (2018) in his research stated that the goal-oriented customers normally have a shopping plan in their minds as determination of customer purchase behaviour. Advertisement with appropriate information will provide information to this group of customers in making purchase decision. Besides that, a good advertisement is beneficial to customer’s purchase decision making, and increasing customer satisfaction and brand loyalty (Chua and Sharma, 2005). Therefore, it is assumed that customers purchase behaviour will be affected by the online advertising more easily if the advertisement matched the customer's favour (Nur et al., 2020). 138 of 225 ICDXA/2021/14 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Product Risk is the perception that a purchased product may not work as originally anticipated (Kim et al., 2008). Although online retailers are available 24/7 with all the product details, nonetheless, it is difficult for customers to analyse the real products based on the images and reviews provided and thus make a wrong judgement over the product quality (Sirkka et al., 1999; Bhatnagar et al., 2000). Although the customers have the higher power in the online shopping experience (evaluation and comparison of product is not affected by salesperson), nevertheless, the whole shopping process has been shifted from tangible to intangible/digital process with high product quality risk (Emad, 2013). This phenomenon has significantly affected customer shopping behaviour (Kok et al., 2020; Lu et al., 2016; Emad, 2013). Delivery Risk is another common factor that happens during online shopping process which issue in product lost, damaged or sent to wrong address (Emad, 2013; Yu et al., 2007; Yeniçeri and Akin 2013; Ankita and Smita, 2015). Customers are also concerned about the slow delivery, improper product packaging and handling during transportation. (Claudia, 2012). Adnan (2014) also advised that online merchants should provide insurance coverage to online buyers to compensate late delivery cases. It is assumed that delivery risk will affect customers’ online shopping behaviour. Security is an important role in the online shopping as it will affect the online shopping decisions of the customers (Emad, 2013). Customers need to fill up the banking details when purchasing online which makes them worried about their personal information getting exposed by the seller to third parties (Rasool et al., 2017). There is research that found that customer attitude, perceived security and perceived trust are among the factors that affect the online shopping behaviour (Meskaran, 2015, Prasetyo et al. 2021). In the light of these research, Online Advertisement, Product Risk, Delivery Risk, and Information Security are selected as independent variables in this research. 3.0 CONCEPTUAL MODEL This research modified Emad (2013) conceptual model by selecting factors that have relationship with online shopping behavior (as shown in Figure 1) with additional famous factor – online advertising. These are the four factors believed to have high influence over college students based on the literature study. Online Advertising Product Risk (PR) Delivery Risk Information (OA) H(3DR) Security (IS) H2 H1 Online Shopping Behavior H4 Figure 1. Research Model – factors affecting college students’ online shopping behavior. The research hypotheses are as following: H1: Online Advertising has significant effect on online shopping behavior among college students. H2: Product Risk has significant effect on online shopping behavior among college students. H3: Delivery Risk has significant effect on online shopping behavior among college students. H4: Information security has significant effect on online shopping behavior among college students. 139 of 225 ICDXA/2021/14 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 4.0 RESEARCH METHODOLOGY Online questionnaire in Google Form is distributed to TAR UC college students and shared in several social media platforms which are Facebook, Instagram and WhatsApp. Simple random sampling method is applied. One-way ANOVA and Cohen’s F effect size calculated using PSPP and F calculator (Zhiyong et al., 2015). The questionnaire is designed in two sections: Section A that focus on demographic (age and gender), and Section B on the Shopping Behaviour, Online Advertising, Product Risk, Delivery Risk and Information Security as shown in Figure 2. These questionnaire items are adopted from several researches in selecting the most suitable and applicable questions for college students. Section B: Emad,2013 Please express your agreement with the following statements on a scale of 1-5 Online Shopping Behavior Q1.Do you think that you can get detailed product information when you shop online? Q2. Do you think that using the Internet for online shopping is easy? Q3. Do you think that online shopping can give facility of easy price comparison? Online Advertising Anupkumar Q4. You often see online advertisements on social media, email and web banners. and Snehal, Q5. I will pay attention to Internet advertising. 2019 Q6. Online advertisements are informative and create awareness. Product Risk Mohammad, Q7. Are you worried about receiving malfunctioning products? 2012 Q8. Are you worried about didn’t receiving what you ordered through online shopping? Q9. Do you think that it is hard to determine the quality of products through the Internet? Delivery Risk Emad, 2013 Q10. You might not receive the product ordered online. Q11. Delivery may be sent to the wrong place. Q12. Sellers may not have timely delivery. Q13. It is not easy to cancel orders when shopping online. Q14. The goods returned may be waiting a long time. Information Security Chong and Q15. Using online shopping can make someone use your username and read your Siti, 2020 transactional information. Q16. Using online shopping can make someone use your username and make orders. Q17. Using online shopping can make someone steal your account information. Figure 2. Questionnaire Items for four independent factors – online advertising, product risk, delivery risk and information security; and one dependant factor – online shopping behavior. 5.0 RESULTS AND DISCUSSION A total sample size of 151 was successfully collected with 53.6% female students and 46.4% male students Based on the Conbach Alpha’s reliability test, the questionnaire items are acceptable at the index of 0.78 (Table 1). 140 of 225 ICDXA/2021/14 @ICDXA2021

International Conference on Digital Transformation and Applications (ICDXA), 25-26 October 2021 Table 1. Reliability test Cronbach Alpha’s Number of items retained .78 17 Table 2. One-Way Anova and Cohen’s f, Effect Size analysis. One-way ANOVA Effect Size, f F Sig. Mean Std Dev f Online Advertising (OA) 2.19 0.026 3.37 0.78 0.57 Product Risk (PR) 2.23 0.023 4.13 0.67 0.43 Delivery Risk (DR) 6.04 0.000 3.65 0.90 0.60 Information Security (IS) 4.36 0.000 3.75 0.78 0.59 Dependent Variable: Shopping Behaviour (SB) Q1, Q2, Q3 Independent Variables: Online Advertising – Q4, Q5, Q6 Product Risk – Q7, Q8, Q9 Delivery Risk – Q10, Q11, Q12, Q13, Q14 Information Security – Q15, Q16, Q17 f = 0.1 = small effects; f = 0.25 = medium effect; f = 0.40 = large effects (Cohen, 1988) Table 2 shows the result of One-way ANOVA test of dependant variable: shopping behavior and independent variables: Online Advertising, Product Risk, Delivery Risk and Information Security. Besides, effect size of each pair OA → SB, PR → SB, DR → SB, IS → SB is calculated in Table 2. The result shows that Online Advertising, F = 2.19 and Sig = 0.026, has a significant effect on shopping behaviour with large effect, f = 0.57 (f > .40 = large effects). Therefore, H1 is supported. Product Risk also has a significant effect on shopping behaviour with F = 2.23 and Sig = 0.023. The effect size of Product Risk on shopping behaviour is f = 0.43 which is a large effect too. Therefore, H2 is supported. Delivery Risk and Information Security also show significant and large effect on the shopping behaviour based on Table 2 (DR: F = 6.04, Sig = 0.000, f = 0.60; IS: F = 4.36, Sig = 0.000, f = 0.59). Therefore, H3 and H4 are supported. Among these factors, Delivery Risk shows the highest effect on shopping behaviour, followed by Information Security, Online Advertising, and lastly Product Risk with the least effects. These results are compatible with all the previous research. The research result clearly stated that Online Advertising does affect college students’ shopping behavior and they agree that these advertisements are informative and helpful in their shopping process (Mean = 3.37). Student is also worry about incorrect product received or no product is received after placing order (Mean = 4.13). Besides, most of the students also agree that Delivery Risk is a concern in their shopping experience (Mean = 3.65). Students nowadays also has high Information Security awareness and they agree with the risk of sharing personal information in the shopping process (Mean = 3.75). Among these factors, Product Risk has the highest concern among students, followed by Online Advertising, Information Security, and Delivery Risk. It is interesting to see that although most of the students are most concerned about Product Risk (Mean = 4.13) but this factor brings the lowest effects in the shopping behavior (f = 141 of 225 ICDXA/2021/14 @ICDXA2021