Fusing local topology and global semantics for link prediction in heterogeneous information networks

Link prediction in Heterogeneous Information Networks (HINs) is both critical and challenging due to the complex interplay between structural sparsity and semantic heterogeneity. While recent subgraph-based methods have achieved success by capturing fine-grained local topology, they often suffer from a limited perspective, failing to infer missing links driven by long-range global semantic dependencies. Conversely, global embedding methods are effective in capturing macro-level semantics but frequently struggle with structural indistinguishability and feature over-smoothing. To address these limitations, we propose SCALP (Structural and Community-Aware Link Prediction), a unified framework that integrates fine-grained local structural information with global semantic consistency. SCALP consists of two mutually reinforcing components: (1) a Local Subgraph Encoder, which extracts enclosing subgraphs around target node pairs and leverages Double-Radius Node Labeling (DRNL) combined with GraphSAGE to accurately capture fine-grained local structural information; and (2) a Multi-view Community Encoder that introduces a differentiable soft clustering mechanism. This mechanism dynamically disentangles latent semantic communities across diverse meta-path views and fuses them via an adaptive attention module to resolve semantic discrepancies. As for the evaluation, we conduct extensive experiments on four public datasets. The results demonstrate that SCALP consistently outperforms state-of-the-art baseline methods.