A Hierarchical Pooling Method Using Structural Entropy and Information Bottleneck

The problem of graph classification in graph neural networks has received extensive research. In the graph classification tasks, graph pooling is one of the important research directions, which aims to reduce the dimensionality of the representation and obtain the prediction label for the entire graph. However, due to fixed compression quotas and node discarding strategies, the commonly adopted hierarchical pooling methods often suffer the graph structure destruction and information loss. And their model training will be time-consuming. To address these issues, we suggest a hierarchical pooling method using structural entropy and information bottleneck (HSEIBPool). In this work, a global optimization algorithm is proposed to convert graph data samples to coding trees and generate cluster assignment matrices in one step. Then, we introduce information bottleneck principle to optimize the expressiveness and robustness of the model, which is applied to process the pooling layer representations. Our model consistently outperforms previous models on various datasets, as demonstrated by the results of our experiments.