Multi-graph learning with adaptive graph-bag mapping.

Area of Science:

• Machine Learning
• Graph Theory
• Data Representation

Background:

• Real-world objects have complex structures difficult to model with single graphs.
• Multi-Graph Learning (MGL) represents objects as bags of graphs, but existing methods limit structural information exploitation.
• Predefined graph-bag mapping assumptions hinder generalization in MGL.

Purpose of the Study:

• To propose a novel Multi-Graph Learning method with Adaptive graph-bag Mapping (MGLAM).
• To fully exploit graph structural information and adaptively model graph-bag mapping.
• To explore different MGL prediction paradigms and their impact.

Main Methods:

• Leveraging graph kernels for initial graph representations to capture structural information.
• Proposing an attention-based multi-graph pooling mechanism for adaptive graph-bag mapping, ensuring permutation invariance.
• Conducting a systematic analysis of various prediction paradigms in MGL.

Main Results:

• MGLAM outperforms state-of-the-art baselines on eight benchmark MGL datasets.
• Achieved average improvements of 3.88% in accuracy, 4.21% in precision, and 2.94% in F1 score.
• Demonstrated significant gains in AUC (0.65%) and reduction in FPR (4.53%).

Conclusions:

• MGLAM effectively exploits graph structure and adaptively learns graph-bag mappings.
• The proposed method offers improved generalization across diverse MGL scenarios.
• MGLAM represents a significant advancement in Multi-Graph Learning performance and applicability.