Multi-graph learning with adaptive graph-bag mapping.

Abstract

Real-world objects usually inherently possess rich and complex structures that are difficult to model using a single graph or instance-based representation. Multi-Graph Learning (MGL) addresses this by representing each object as a bag of graphs, each encoding distinct local structures. While this formulation enhances expressiveness, existing methods typically convert graphs into binary vectors, limiting their ability to exploit the structural information inherent in graphs. Meanwhile, they often rely on predefined graph-bag mapping assumptions, which constrain them to fixed graph-bag relationships and hinder generalization across diverse MGL scenarios. Moreover, the differences and impacts of various prediction paradigms have not been fully explored. To address these challenges, we propose a novel Multi-Graph Learning method with Adaptive graph-bag Mapping (MGLAM). MGLAM fully exploits graph structural information, adaptively models the graph-bag mapping without relying on predefined assumptions that constrain fixed graph-bag relationships, and thoroughly explores different MGL prediction paradigms. Specifically, we first leverage graph kernels to construct initial graph representations, facilitating effective capture of graph structural information. Then, we demonstrate that graphs within a bag must satisfy permutation invariance, and propose an attention-based multi-graph pooling mechanism to adaptively learn the graph-bag mapping without relying on predefined assumptions. We further conduct a systematic analysis of different prediction paradigms and their impact on MGL performance. Experimental results on eight benchmark MGL datasets show that MGLAM outperforms state-of-the-art baselines, with average improvements of 3.88 %, 4.21 %, 2.94 %, 0.65 %, and 4.53 % in accuracy, precision, F1 score, AUC, and FPR, respectively.