Enhancing Generalization in Graph Neural Networks: Challenges and Solutions

Speaker: Yujun Yan (Dartmouth)

Date: 11/8/24

Abstract: As deep learning continues to expand, graph neural networks (GNNs) have emerged as powerful tools for processing complex graph data. Despite their promise, GNNs face significant challenges with generalization at multiple levels. At the node level, they often prioritize nodes with higher degrees and stronger homophily, resulting in skewed performance that favors these nodes. At the graph level, GNNs struggle to generalize to larger, unseen graphs, limiting their adaptability to new data scales. Furthermore, the similarity relationships among graphs, as captured by GNNs, may not remain consistent across layers, further hindering their ability to generalize effectively. This talk explores these critical generalization challenges and presents strategic solutions to improve GNN performance across diverse graph structures.