MGLP: Multi-Granularity Position Embedding for Link Prediction

A new research paper on arXiv (2607.29115) introduces MGLP, a method for link prediction in graphs that leverages multi-granularity position embeddings. The approach addresses limitations of previous methods that used single-granularity landmarks, which neglect the hierarchical nature of homophilic structures. MGLP employs an Adaptive Granular-Ball Graph Refinement mechanism to adaptively partition the graph into homophilic subdomains at optimal granularity levels, with central nodes within these subdomains serving as landmarks for position encoding. This allows for more accurate capture of structural patterns and implicit connections, improving link prediction performance. The paper is categorized as a cross-type announcement and is available on arXiv.