Mutual Information with Experts: An Approach to Graph Pooling for Graph Neural Networks

Graph neural networks are powerful for graph-structured data, with graph pooling critical for tasks like classification. Graph pooling methods discard nodes to form a coarser graph representation, while recent expert pooling approaches leverage multiple experts to score and robustly select nodes for retention, capturing diverse graph characteristics, leading to poor performance on complex graphs. To address this, we propose Mutual Information Graph Pooling with Experts, which first assigns each expert to evaluate only a subset of similar nodes, improving scalability and performance on complex graphs, and, second, incorporates frequency-domain representations, ensuring applicability across diverse graph types. We perform pooling based on mutual information across both frequency and spatial domains. Extensive experiments demonstrate that our method outperforms state-of-the-art pooling techniques.