Robust link prediction in heterogeneous graphs via diffusion-based generative fusion and relation-aware guidance

Link prediction in heterogeneous graphs, which captures both topological and semantic information in graph structures helps the discovery of meaningful relationships. However, several fundamental challenges of link prediction remain insufficiently addressed. First, real-world data are noisy, causing existing models to struggle with learning robust representations. Second, although fusing information from multiple relation types is a core component of heterogeneous graph learning, most existing methods rely on simple aggregation methods such as a weighted sum. Such naive fusion may lead to embeddings that stray from their underlying manifold structure, ultimately degrading the accuracy of link prediction. Moreover, each link type is not equally informative. To address these challenges, we propose a novel diffusion-based framework for heterogeneous graph link prediction (DiffHGLP). Our proposed framework leverages a diffusion model to jointly denoise and fuse multi-relational information in heterogeneous graphs and employs an attention-based guidance mechanism for selecting links with informative link types. Extensive experiments on three public benchmark datasets show that DiffHGLP consistently outperforms existing state-of-the-art methods.