LDSPool: Latent Dirichlet structure pooling with hierarchical graph context representation

In graph data analysis, particularly the graph classification task, a discriminative graph-level representation is significant to improve classification performance. For the performance improvement, recent studies have applied pooling methods to graph neural networks. However, the existing pooling approaches lose the initial graph structural information when incorporating each node. When a latent structure obtained from the pooling operation is given, the nodes in each latent structure have a different significance compared with the original graph. This structural information discrepancy between initial and latent structures leads to an inadequate graph representation when the existing methods generate the graph result. Motivated by this, we propose latent Dirichlet structure pooling (LDSPool) with a hierarchical graph context representation. After independently learning both structures, the proposed LDSPool incorporates the original graph and latent structure to reduce the structural discrepancy. We employ latent Dirichlet allocation (LDA) to extract the latent structure from the original graph. To validate the performance of LDSPool, we benchmarked five public graph datasets and seven existing pooling methods. The experimental results and ablation studies demonstrate that LDSPool achieves better performance than previous methods.