GSGSL: Gravity-driven self-supervised graph structure learning

Abstract

The existing graph structure learning methods heavily rely on the original graph structure and often fail to capture potential high-level abstract features and global correlations within the graph. To address these issues, this paper proposes a Universal Gravitational-driven Self-supervised Graph Structure Learning method (GSGSL), overcoming the limitations of current graph structure learning methods in advanced feature extraction. GSGSL models the graph structure using the universal force of gravity in a dynamic, globally adaptive manner. Additionally, it employs a multi-perspective contrastive learning approach to eliminate the need for external labels, jointly optimizing graph structure learning with downstream tasks. Extensive experimental results on public datasets demonstrate that, in comparative experiments without considering the original adjacency matrix, the GSGSL method outperforms baseline models by 0.5% to 17.3%. In comparative experiments optimizing the original adjacency matrix, the GSGSL method exhibits improvements ranging from 0.9% to 12%, validating that the GSGSL approach to simulating gravitational fields has better dynamic characteristics. It effectively captures advanced abstract features and global characteristics of graph data, surpassing the limitations of baseline methods in graph structure learning.