How Useful Is Graph Pooling for Node-Level Tasks?

Abstract

As an essential component of graph neural networks, graph pooling is indispensable for graph-level tasks such as graph classification and generation. However, certain node-level tasks inherently require graph pooling, particularly multiple instance learning (MIL) on graphs , a weakly supervised learning paradigm where only set-level labels are available for training node-level predictors. Existing embedding-based pooling aggregates node embeddings to obtain a holistic graph-level representation, neglecting direct inference of node labels. To address this limitation, we propose instance-based pooling , which maps node embeddings to node probabilities before generating graph representations. We prove that embedding-based pooling methods can be seamlessly transformed into instance-based ones without losing permutation invariance or expressiveness, while the latter offers better interpretability. Extensive experiments on diverse benchmark datasets validate the effectiveness of our proposed method, providing key insights into the selection of pooling methods for different machine learning tasks on graphs.