School of Engineering Department of Computer Science and Engineering 100 AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: CHIU Ka Ho / COSC Course: UROP3100, Spring WiFi fingerprinting can be of great importance in indoor localization, monitoring objects, and tracking people, which gives value to the study of the method to analyze it. In the previous work during UROP 2100, we have developed a model that is based on network embedding and clustering method, and have performed some experiments on the data collected from the second floor of the main building of HKUST. In this report, we will cover some improvements that we have made to the existing model and would provide more comparisons to illustrate the performance of our model. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: PENG Zhuoxuan / COSC Course: UROP3100, Spring License plate recognition (LPR) is a deeply explored task, but difficult to be implemented on compact computers such as Raspberry Pi. Last semester I implemented a calibration algorithm which can recognize all the areas where a license plate may appear in a fixed scene. To further reduce the computation and accelerate the recognition, it is necessary to compress the recognition model based on neural network. Therefore, I attempted several different model compression techniques and evaluated their effectiveness. Results show that it is possible to accelerate the recognition process by 3 times without much loss of accuracy. In the next step I will focus on the implementation of the IoT system. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: XIANG Letian / COMP Course: UROP3100, Spring In this semester, I continue to work on the traffic forecasting problem using spatial-temporal data analysis. In traffic forecasting, we aim to predict the inflow and outflow of traffic of a region in the coming time interval using the historical spatiotemporal data. While existing methods may not be light and have limited efficiency, we would like to propose a new model called Spatial-temporal Dynamic Context Model (ST-DCM) for traffic forecasting. ST-DCM makes use of context and self-attention. It is evaluated to outperformed the state-of-the-art models in terms of accuracy and efficiency.