HiCAP: Hierarchical Clustering-based Attention Pooling for Graph Representation Learning

Graph representation learning has gained increasing importance in the analysis of complex relational data. The utilization of Graph Neural Networks (GNNs) has demonstrated significant potential for capturing graph structures and node features in this domain. However, the challenge of graph pooling, which involves summarizing graphs into concise representations, remains a persistent obstacle. In this study, a novel method called HiCAP (Hierarchical Cluster-Based Attention Pooling) is proposed to address the limitations of existing graph pooling approaches. By combining the strengths of node clustering and node dropping techniques, HiCAP establishes a hierarchical framework. Initially, a soft cluster assignment matrix is learned through the application of a GNN. Subsequently, the matrix undergoes a transformation into a hard assignment by incorporating structural considerations. Attention-based scoring is subsequently employed to select representative nodes within each cluster. Extensive experimental evaluations on benchmark datasets showcase the effectiveness of HiCAP in graph classification tasks, outperforming state-of-the-art baselines. This method introduces an innovative approach to graph representation learning, thereby advancing graph pooling techniques.