DiffMBR: diffusion-guided GCN for multi-behavior recommendation

Current multi-behavior recommendation methods are often sensitive to noise and sparsity in auxiliary behaviors and comprehensively model users’ intrinsic behavioral habits. To address these limitations, we propose a diffusion-guided graph convolutional framework for robust multi-behavior recommendation, named DiffMBR. DiffMBR first introduces a dual-level structural diffusion mechanism to denoise and enhance both behavior-specific and unified interaction graphs. This process recovers reliable latent connections while suppressing noisy signals. Building on the refined graph structures, a dual-path propagation architecture is then used to capture personalized behavioral dependencies via cascade propagation, and behavior-invariant global preferences via consensus diffusion propagation. These complementary representations are further aligned through cross-view contrastive learning to obtain a unified and robust embedding. We conduct extensive experiments on three real-world multi-behavior datasets, Tmall, Taobao, and Beibei, showing that DiffMBR consistently outperforms various state-of-the-art methods. Notably, on the Taobao dataset, our model achieves relative improvements of 12.36% in Hit@10 and 25.31% in NDCG@10 over the strongest baseline. Additional experiments further illustrate the robustness and effectiveness of each component.