MGLP paper introduces granular-ball position embedding for graph link prediction
New graph embedding method captures hierarchical structure to improve link prediction accuracy
Link prediction aims to uncover potential or future connections in a graph, and position information is key to distinguishing nodes by their relative relationships. The proposed MGLP method introduces an Adaptive Granular-Ball Graph Refinement mechanism to adaptively split graphs into homophilic subdomains, whose central nodes form a Hierarchical Central Graph. A novel multi-granularity hierarchical distance encoding then captures both homophilic structures and their hierarchical correlations, boosting discriminative power. Experiments show the method performs excellently and is highly competitive against baseline algorithms, and the code is available via the provided link.
- MGLP introduces Adaptive Granular-Ball Graph Refinement to partition graphs into homophilic subdomains at optimal granularity levels
- Central nodes in subdomains form a Hierarchical Central Graph, enabling multi-scale positional representation
- The method outperforms baseline link prediction algorithms, with code released on GitHub
Why It Matters
Improves link prediction precision in social networks and knowledge graphs by leveraging hierarchical structural information.