MEDKGC: A Mixture-of-Experts Enhanced Diffusion Model for Multimodal Knowledge Graph Completion

Multimodal Knowledge Graphs (MKGs) find broad applications across various downstream tasks, such as recommendation systems and information retrieval. However, these knowledge graphs often suffer from inherent incompleteness, which has motivated the development of Multimodal Knowledge Graph Completion (MKGC) techniques. Most MKGC methods faces significant challenges in achieving fine-grained multimodal information capturing in the knowledge graphs. To address these limitations, we propose Mixture of Experts Enhanced Diffusion Model for Multimodal Knowledge Graph Completion (MEDKGC), a novel framework that fundamentally enhances diffusion models through Mixture-of-Experts(MoE) for MKGC. The framework introduces two key innovations: a Relation-Aware Mixture-of-Experts (RAMoE) module that provides fine-grained, relation-specific condition, and a Multimodal Context Diffusion Model (MCDM) that leverages this guidance for accurate generative reasoning. Then, through a carefully designed two-stage training strategy, our framework effectively coordinates the dynamic feature learning of MoE with the distribution modeling capability of diffusion processes. Extensive experiments demonstrate that MEDKGC achieves state-of-the-art performance across multiple benchmarks, with significant improvements over existing discriminative and generative methods. The framework exhibits remarkable adaptability to diverse relation types and modality requirements, establishing a new paradigm for generative knowledge graph completion.