AutoGSP: Automated graph-level representation learning via subgraph detection and propagation deceleration

Graph neural networks have demonstrated impressive performance across various learning tasks involving graph-structured data. However, effectively capturing and leveraging the inherent and widely recognized underlying substructures in graphs to enhance graph-level learning performance remains a significant challenge. Existing methods rely on hand-crafted heuristics and domain-specific knowledge, which limits their effectiveness, flexibility, and generalizability across diverse datasets and tasks. The over-smoothing in smallscale subgraph structures hinders the representational capacity of graph-level learning. In this paper, we propose a novel automated graph-level representation learning framework, namely AutoGSP, to learn the informative topological properties of graphs. To capture the dominant critical subgraph structures for specific tasks, AutoGSP adopts reinforcement learning for automated subgraph detection. To exploit abundant information in critical subgraph structures, AutoGSP enhances their representation by utilizing the propagation deceleration mechanism on small-scale subgraphs. These two modules are trained via iterative optimization to provide the critical subgraph structures as the abstract representation of the entire graph without requiring prior knowledge. Extensive experiments on five real-world benchmark datasets (MUTAG, PTC, PROTEINS, IMDB-B, and IMDB-M) demonstrate the effectiveness and generalizability of our proposed methods. Multiple ablation studies validate the contributions of each component within our method. AutoGSP enables significant practical applications on graph-level learning tasks, including molecular property prediction and social network analysis.