Applied Math Seminar- Jianheng Tang, Hong Kong University of Science and Technology

Abstract: Similarity computation and anomaly detection are fundamental data mining tasks, but the heterogeneous, relational-centric, and non-Euclidean nature of graphs presents unique challenges. This talk explores novel approaches to both problems in the context of graph data. The first part will focus on graph similarity, addressing both node similarity across attributed networks and graph similarity within databases. Leveraging optimal transport and the Gromov-Wasserstein distance, we demonstrate promising results in applications such as knowledge graph entity alignment and social network account linkage. The second part will discuss my work on graph anomaly detection, answering three key questions: (1) which type of graph neural networks are most effective for graph anomaly detection? (2) are industry-favored decision tree ensembles superior to graph neural networks for this task? and (3) how can we unify the detection of different types of graph anomalies within a single framework? Finally, the talk will share some insights on the emerging intersection of large language models and graph machine learning.