Title: Providing Explanations for Unsupervised Graph Learning Models

Abstract: This talk presents advancements in providing explanations for unsupervised graph learning models. It highlights eXplainable AI (XAI)'s role in identifying influential subgraphs for graph learning models. A key contribution, the UNR-Explainer (ICLR 2024) , generates counterfactual explanations for unsupervised node representation learning models by using Monte Carlo Tree Search to find important subgraphs. Additionally, the talk introduces the HINT-G framework (upcoming KDD 2025) , which leverages influence functions to explain GNNs across supervised and unsupervised settings, uniquely considering the impact of non-existent edges. HINT-G demonstrates superior performance in identifying both positive and negative influential edges, proving robust and task-agnostic. Its theoretical underpinnings reveal that high influence correlates with class probability flips in supervised models and non-homophilic edges in unsupervised models, allowing for efficient influence acquisition and even serving as an early stopping indicator for unsupervised GNN training.

Bio: Hogun Park is an Associate Professor in the Department of Computer Science and Engineering at Sungkyunkwan University (SKKU), South Korea, where he also serves as head of the Intelligent Software graduate program. His research centers on machine learning and data mining, with a particular focus on self-supervised graph learning, graph neural networks, knowledge graphs, and explainable AI. He earned his Ph.D. in Computer Science from Purdue University under the supervision of Prof. Jennifer Neville, with a dissertation on explainable graph neural networks for semi-supervised learning. His work has been widely published in top-tier venues such as KDD, WWW, ICML, ICLR, AAAI, and ACL. He is a Fulbright Fellowship recipient, and his research has been selected for multiple oral presentations at premier conferences, including AAAI and WSDM.