D3MBR: Dual-Level Diffusion Denoiser with Preference Guidance for Multi-Behavior Recommendation
TL;DR
This work reshape the denoising problem as a clean data generation task, and proposes a noise-free generation paradigm in a data-centric perspective, and integrates D \({}^{3}\) MBR with advanced MBR models and achieves great performance improvements across four datasets.
Abstract
Multi-Behavior Recommendation (MBR) integrates rich auxiliary behaviors to enhance user preference modeling in target behaviors, garnering wide research attention. Existing efforts focus on improving the modeling capability of MBR via sophisticated models based on fixed datasets, which are essentially model-centric. However, they overlook the potential data quality issues, such as noise. To bridge this gap, we reshape the denoising problem as a clean data generation task, and propose a noise-free generation paradigm in a data-centric perspective. Based on this paradigm, driven by the generation and denoising capabilities of the Diffusion Models, we devise a Dual-level Diffusion Denoiser with preference guidance (D \({}^{3}\) MBR). D \({}^{3}\) MBR tailors to the distinct noise characteristics of target and auxiliary behaviors. Specifically, for the target behavior, we propose a representation-level denoising strategy that utilizes collaborative preference signals to guide a diffusion denoiser in generating noise-free representations. For auxiliary behaviors, we design an interaction-level denoising strategy that uses denoised user preferences from the target behavior to direct another diffusion denoiser in generating clean interactions. To validate the effectiveness, we integrate D \({}^{3}\) MBR with advanced MBR models and achieve great performance improvements across four datasets. These results highlight the potential of our method in advancing MBR by addressing data quality issues.