SSPool: A Simple Siamese Framework for Graph Infomax Pooling

• Published in: IEEE transactions on network science and engineering 2024-01, Vol.11 (1), p.1-9
• Main Authors: Li, Zhi-Peng, Su, Hai-Long, Zhu, Xiao-Bo, Gribova, Valeriya, Filaretov, Vladimir Fedorovich, Huang, De-Shuang
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Coarsening; Computational modeling; Graph neural networks; graph pooling; Graph representation learning; Graphical representations; Mutual information; mutual information maximization; Representation learning; simple Siamese networks; Spectral analysis; Task analysis; Topology
• ISSN: 2327-4697; 2334-329X
• DOI: 10.1109/TNSE.2023.3300878

Graph pooling plays an indispensable role in graph representation learning by aggregating the node representations on the graph into a compact form. However, the existing graph pooling models based on mutual information maximization all need to construct negative samples and usually only consider local neighborhood information. In this paper, we propose a novel coarsening graph infomax pooling method called SSPool, which selects the most informative subset of nodes as the pooled graph by neural estimation of the mutual information between the coarsened graph and the augmented coarsening graph. To maximize the neural estimation of mutual information, we propose a self-attention-based top-k method to learn the coarsened graph, while we use edge perturbation to augment the coarsened graph and a simple Siamese network framework to help the coarsened graph to get a better representation. Extensive experimental results show that our method outperforms previous state-of-the-art methods on eight benchmark data sets.