School of Engineering Department of Computer Science and Engineering 119 Graph Machine Learning for Logical Reasoning Supervisor: SONG Yangqiu / CSE Student: LI Yifan / COMP Course: UROP 1100, Spring This progress report presents an ongoing UROP project that explores the application of complex query answering (CQA) using graph neural networks (GNN) on knowledge graphs (KG). The report introduces the concepts of logical reasoning, knowledge graphs, and GNN, and extensively reviews relevant literature on representation learning techniques and query-answering methods for KGs. The research project includes a novel model of CQA under development, and several experiments on training techniques are reported. The discussion analyzes current experimental results, limitations encountered, and potential future directions. This research provides insights into the role of GNN in complex query answering on KGs, laying the foundation for further advancements in the field. Graph Machine Learning for Logical Reasoning Supervisor: SONG Yangqiu / CSE Student: TSANG Hong Ting / COMP Course: UROP 1100, Fall UROP 2100, Spring This project aims to merge logical reasoning and machine learning to solve the problem of complex query answering on knowledge hypergraphs. In the LMPNN (Logical Message Passing Neural Network) approach [Wang et al., 2023], a message passing network incorporating pretrained knowledge graph representations achieved state-of-the-art results on the complex query answering task. Furthermore, in recent research [Huang et al., 2024], a new framework for relational hyperedge message passing was proposed. Our project defined hyperedge query answering, created a new dataset for complex query answering in Hyper Knowledge Graph, and proposed methods for complex query answering on hyperedge graphs using a pretrained neural hyperedge predictor combined with a hyperedge message passing framework. Our code provides a pipeline to create datasets from any hypergraphs. Graph Machine Learning Methods for Scientific Discovery Supervisor: SONG Yangqiu / CSE Student: CHAN Cheuk Hang Marcus / COMP Course: UROP 1100, Spring Biological and chemical fields has been widely using DBSCAN to denoise their images. However, while the accuracy is satisfactory, researchers have to manually tune the hyperparameter: elbow threshold so as to filter the noises. This creates inconveniences to them and poses challenges to beginners. This problem leads to an idea of using machine learning to automate the process. Due to the irregularity in the molecules’ shape, and especially the permutation equivariance property, graph machine learning is the best fit. This report presents a method that leverages the Graph Substructure Network using identifiers and point distances as the feature vector.
RkJQdWJsaXNoZXIy NDk5Njg=