DMSPool: Dual Multi-Scale Pooling for Graph Representation Learning

• Published in: Database Systems for Advanced Applications Vol. 12681; pp. 375 - 384
• Main Authors: Yu, Hualei, Luo, Chong, Du, Yuntao, Cheng, Hao, Cao, Meng, Wang, Chongjun
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Graph convolution; Graph neural networks; Graph pooling; Multiple GNN architectures
• ISBN: 3030731936; 9783030731939
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-73194-6_25

Graph neural networks (GNNs) have recently become a powerful graph representation technique for graph-related tasks. However, the existing GNN models mainly focus on generalizing convolution and pooling operations in a pre-defined unified architecture, limiting the model’s ability to capture meaningful information of nodes or local structures. Besides, the importance of subgraphs at various levels has not been well-reflected. To address the above challenges, we propose Dual Multi-Scale Pooling (DMSPool), which uses multiple architectures concurrently to integrate graph convolution and pooling modules in an end-to-end fashion. Specifically, these modules adopt multiple GNN architectures to learn node-level embeddings and nodes’ importance from different aggregation iterations. Additionally, we employ attention mechanism to adaptively determine the contribution of subgraphs’ representations at varying levels to graph classification and integrate them to perform the cross-scale graph level representation. Experiment results show that DMSPool achieves superior graph classification performance over the state-of-the-art graph representation learning methods.