School of Engineering Department of Computer Science and Engineering 148 Knowledge Discovery over Database Supervisor: WONG Raymond Chi Wing / CSE Student: LUI Ka Kit / COMP Course: UROP 4100, Spring Shortest path query has been one of the most important problems in graphs. In data mining and graph databases, shortest path is a fundamental operation. In road networks, the graph weights, which represents time to travel through the roads, are seldom static, and may change over time. For example, a road network may become congested during rush hours, and it takes longer time to travel through a road. In this report, we study dynamic road networks and finding shortest path on such networks, and when a user should depart so that the shortest path can be travelled and reach the destination in the shortest amount of time. Knowledge Discovery over Database Supervisor: WONG Raymond Chi Wing / CSE Student: NIE Weihao / COMP Course: UROP 3100, Fall UROP 4100, Spring Recommendation systems are critical in both e-commerce and online recruitment platforms, where they help users navigate vast amounts of information. While job recommendation systems prioritize aligning candidates with job opportunities based on nuanced and dynamic preferences, product recommendation systems focus on matching users with items they are likely to purchase. This report explores the feasibility of applying advanced product recommendation methods, including LLM-based, session-based, and graphbased approaches, to the field of job recommendation. By analyzing several recently proposed methods, this report identifies potential directions for improvement in job recommendation systems. Additionally, due to the scarcity of publicly available job recommendation datasets, this report introduces the Zhaopin dataset, a publicly accessible dataset. Details about the dataset, including its characteristics and the preprocessing steps performed, are also presented in this report. Knowledge Discovery over Database Supervisor: WONG Raymond Chi Wing / CSE Student: YIP Sau Lai / DSCT Course: UROP 2100, Fall UROP 3100, Spring This UROP project continues the development of an LLM-based VQL debugger, which consists of five key components: Corrector, Retriever, Detector, Schema Encoder, and Action Planner. A refined experiment was conducted on the Corrector to address issues including insufficient erroneous VQL examples and inappropriate data splitting. The results indicate that the VQL PyDict approach outperforms the SQL PyDict + VisBin PyDict approach. Further analysis reveals that information consistency is more critical than representation simplicity for Code-LLM in this debugging task. Additionally, the Retriever component was implemented with a sequential architecture. This architecture first retrieves examples using a GNN-based encoder to model similarity in VQL, followed by narrowing the selection based on the similarity of natural language queries using a pre-trained text encoder.
RkJQdWJsaXNoZXIy NDk5Njg=