일시: 2025. 3. 27.(목) 2:00pm ~ 3:00pm
장소: 제4공학관 D508

Presentor: Dr. Dongwoo Kim (김동우) / Associate Professor at the Graduate School of Artificial Intelligence, Pohang University of Science and Technology (POSTECH)

Title: Advances in Graph Neural Networks: Bridging Generalization, Expressivity, and Influence Analysis

Abstract: Graph Neural Networks (GNNs) have demonstrated outstanding capabilities in handling complex graph-structured data, yet several fundamental challenges persist, including balancing generalization with expressivity, addressing over-smoothing in heterophilic graphs, and accurately assessing node and edge influence in predictive tasks. This seminar integrates recent advancements addressing these core issues: First, we present techniques mitigating over-smoothing through the introduction of reverse process mechanisms specifically designed for heterophilic graphs, where conventional methods typically struggle. Second, we explore novel approaches for bridging the gap between generalization and expressivity in GNNs, highlighting structural adjustments that preserve representational richness while ensuring robust generalization to unseen data. Lastly, we delve into enhanced influence function methodologies tailored for transductive node classification, offering precise, practical frameworks to quantify how individual graph components impact model behavior and predictive outcomes. We emphasize an innovative formulation that not only accommodates the intrinsic non-convexity of contemporary GNN architectures but also systematically accounts for direct effects of graph modifications such as edge re-wiring. Together, these interconnected studies offer comprehensive insights and advanced tools for improving the design, interpretation, and performance of GNNs in complex real-world scenarios.