TSD-Rec: Metric semantic noise enhanced diffusion based contrastive learning with topology prior for multi-behavior recommendation

Contrastive learning helps alleviate data sparsity by extracting self-supervised signals from raw behavior information, enabling multi-behavior recommendation to more comprehensively characterize user interests. However, most existing contrastive learning-based multi-behavior recommendation models typically generate augmented views of raw behaviors through operations such as random perturbations, which disrupt the latent semantic information of the original behavior views. Moreover, while most of them are explicitly modeled through graph convolutions, they remain plagued by issues such as noise, which can easily lead to structural distortions in graph embedding representations. To address these challenges, we propose TSD-Rec, a novel diffusion-based contrastive learning framework for multi-behavior recommendation. Specifically, we first design a novel metric learning-based behavior similarity diffusion modeling process. This process utilizes metric learning methods to handle the similarity between behaviors, providing a deep understanding of the semantic intensity of different behaviors and ultimately generating semantically consistent behavior noise. The generated behavior noise is then injected into the diffusion process, followed by denoising to generate augmented behavior views, ensuring that the latent semantics of the augmented views remain consistent. Additionally, we design a behavior graph topology pre-training model as a supplement to the user behavior graph model to alleviate noise interference and structural distortion in multi-behavior recommendation, thereby improving recommendation performance. Experimental results on real-world datasets demonstrate that TSD-Rec achieves stable and competitive improvements over baselines on HR@K and NDCG@K.