Adaptive Subgraph Neural Network With Reinforced Critical Structure Mining.

Area of Science:

• Graph Representation Learning
• Machine Learning
• Data Mining

Background:

• Graph representation learning methods are successful but lack transparency regarding exploited knowledge.
• Identifying critical subgraphs is crucial for understanding prediction drivers in graph data.

Purpose of the Study:

• To propose AdaSNN, an Adaptive Subgraph Neural Network, for detecting dominant subgraphs in graph data.
• To enhance subgraph representations using a Bi-Level Mutual Information Enhancement Mechanism for global and label awareness.
• To improve the interpretability of graph mining results by identifying intrinsic graph properties.

Main Methods:

• Developed a Reinforced Subgraph Detection Module for adaptive subgraph searching without predefined rules.
• Implemented a Bi-Level Mutual Information Enhancement Mechanism for maximizing global-aware and label-aware mutual information.
• Evaluated AdaSNN on seven diverse graph datasets.

Main Results:

• AdaSNN achieved significant and consistent performance improvements across multiple graph datasets.
• The method successfully identified critical subgraphs, enhancing prediction accuracy.
• AdaSNN provided insightful and interpretable results by mining intrinsic graph properties.

Conclusions:

• AdaSNN effectively mines critical subgraphs, leading to superior performance in graph representation learning.
• The adaptive subgraph detection and mutual information enhancement contribute to improved accuracy and interpretability.
• This approach offers a promising direction for transparent and effective graph mining.