An enhancement of multi-scope topological graph pooling and representation learning with attention for molecular...

In recent years, recent advancements have shown the successful integration of topological data analysis (TDA) with deep learning (DL) to enhance representations of complex data structures, such as graphs. Graph neural networks (GNNs) have emerged as a powerful tool for analyzing graph-based data and have been popularly applied to multiple tasks related to node and graph analysis and classification. Given the intrinsic topological nature of graph connectivity, recent studies have leveraged topological features; including persistent homology and landmark extraction, to enrich graph representations. These topology-enhanced GNNs have demonstrated significant promise in improving performance across various graph learning problems, such as classification. One key area of innovation is topological graph pooling strategies, which aim to capture multi-scale (both local and global) topology-aware graph-level representations. However, most existing approaches rely heavily on local-neighborhood feature aggregation, making them insufficient for preserving multi-scale representations of complex-structured graphs. These limitations have hindered their ability to efficiently capture the rich topological and structural diversity of input graphs during the pooling process. To address this gap, in this paper, we propose a novel attention-enhanced, multi-scope topological graph pooling strategy, called AETP. Our proposed AETP is designed to extract both discriminative topology-structured information and graph-level variations, enabling high-expressive representation learning. Our AETP model specifically targets complex and small molecular graph representation learning and classification tasks. Comprehensive comparative studies on various real-world molecular datasets, including FDA_DILIst, T3DB_Toxin_2, Eye_Irritation and Eye_Corrosion, validate the effectiveness of our proposed AETP model, demonstrating its superior performance compared to existing graph embedding baselines, such as: GCN, GraphSAGE, GAT, GIN, GINE, UniMP, GATv2, TOGL and TopoPool.