On Exploring Node-feature and Graph-structure Diversities for Node Drop Graph Pooling

Chuang Liu, Yibing Zhan, Baosheng Yu, Liu Liu, Bo Du, Wenbin Hu, Tongliang Liu

Keywords: Graph Neural Networks, Graph Pooling, Graph Classification

Abstract: Graph pooling is essential in learning effective graph-level representations. One mainstream type of graph pooling is node drop pooling, which preserves the nodes in graphs with top-k calculated significance scores. However, despite being commonly adopted, current node drop pooling methods generally ignore node diversity from the perspectives of node features and graph structures. Therefore, they still obtain graph-level representations suboptimally. To address the issue mentioned above, we propose a novel plug-and-play scheme, termed MID, using a \textbf{M}ultidimensional score space with two score operations, \textit{i.e.}, fl\textbf{I}pscore and \textbf{D}ropscore, to explore the node-feature and graph-structure diversities in graphs. Specifically, the multidimensional score space depicts the significance of nodes through multiple criteria; the flipsscore encourages the maintenance of dissimilar features, thus preserving the node-feature diversity; and the dropscore forces the model to notice diverse graph structures instead of being stuck in significant local structures. What is more, we evaluate our proposed MID by applying it to a variety of popular node drop pooling methods, including TopKPool, SAGPool, GSAPool, and ASAP. Extensive experiments on seventeen real-world graph classification datasets demonstrate that our proposed scheme efficiently and consistently brings over 2.8\% improvements in average when using different backbone models and datasets. The datasets include FRANKENSTEIN, IMDB-B, IMDB-M, REDDIT-B, COLLAB from the social domain and D\&D, PROTEINS, NCI1, MUTAG, PTC-MR, NCI109, ENZYMES, MUTAGENICITY, HIV, BBBP, TOXCAST, TOX21 from the biochemical domain.\footnote{Code will be made publicly available at~\url{http://github.com/xxx/xxx}.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2306.12726/code)