DiffMLP: A diffusion-based multi-hop link prediction framework in knowledge graphs.

Abstract

Multi-hop link prediction in knowledge graphs remains a challenging task, particularly in modeling complex reasoning paths and handling uncertainty. Existing approaches often fail to capture the interactive dependencies between the reasoning context and the neighborhood space. To solve this problem, this paper introduces DiffMLP, a novel framework that models the action space at each hop as a conditional distribution learned via a reverse diffusion process. DiffMLP leverages a graph attention-based denoiser guided by prior reasoning context, allowing the model to identify the most relevant action through progressive refinement of action space embeddings. It further incorporates normalized noise injection to stabilize the forward process and utilizes priori constraints to regulate the reverse process. Extensive experiments demonstrate that DiffMLP achieves state-of-the-art results across four benchmarks, notably improving MRR by 7.0 % and Hits@3 by 12.7 % over the previous best model on the FB15K-237 dataset.