Hierarchical Graph Pooling with Structure Learning

TL;DR

The paper tackles the lack of effective pooling for hierarchical graph representations in graph neural networks by introducing HGP-SL, a non-parametric pooling operator paired with a sparse, learnable structure refinement layer. This combination preserves essential topological information while progressively coarsening graphs, enabling richer graph-level representations for classification. Extensive experiments on six datasets show consistent improvements over kernel-based, GNN-based, and pooling-based baselines, with ablations confirming the value of the structure-learning component. The approach is scalable, end-to-end trainable, and the learned sparse graphs offer clearer interpretability of subgraph relationships, boosting practical applicability in real-world graph classification tasks.

Abstract

Graph Neural Networks (GNNs), which generalize deep neural networks to graph-structured data, have drawn considerable attention and achieved state-of-the-art performance in numerous graph related tasks. However, existing GNN models mainly focus on designing graph convolution operations. The graph pooling (or downsampling) operations, that play an important role in learning hierarchical representations, are usually overlooked. In this paper, we propose a novel graph pooling operator, called Hierarchical Graph Pooling with Structure Learning (HGP-SL), which can be integrated into various graph neural network architectures. HGP-SL incorporates graph pooling and structure learning into a unified module to generate hierarchical representations of graphs. More specifically, the graph pooling operation adaptively selects a subset of nodes to form an induced subgraph for the subsequent layers. To preserve the integrity of graph's topological information, we further introduce a structure learning mechanism to learn a refined graph structure for the pooled graph at each layer. By combining HGP-SL operator with graph neural networks, we perform graph level representation learning with focus on graph classification task. Experimental results on six widely used benchmarks demonstrate the effectiveness of our proposed model.

Figures (3)

• Figure 1: Architecture of proposed HGP-SL operator combined with graph neural network. The dashed box demonstrates the workflow of HGP-SL, which involves graph pooling and structure learning. The learned edges are represented as dashed lines in the graph. This procedure (convolution and pooling operations) is repeated several times. Then, a readout function is applied to aggregate node representations to make a fixed size representation, which goes through MLP layers for graph classification.
• Figure 2: Hyper-parameter sensitivity analysis.
• Figure 3: Visualization of different pooling methods.