Recommendation via Collaborative Diffusion Generative Model

Joojo Walker, Ting Zhong, Fengli Zhang, Qiang Gao, Fan Zhou

Abstract: Despite the success of classical collaborative filtering (CF) methods in the recommendation systems domain, we point out two issues that essentially limit this class of models. Firstly, most classical CF models predominantly yield weak collaborative signals, which makes them deliver suboptimal recommendation performance. Secondly, most classical CF models produce unsatisfactory latent representations resulting in poor model generalization and performance. To address these limitations, this paper presents the Collaborative Diffusion Generative Model (CODIGEM), the first-ever denoising diffusion probabilistic model (DDPM)-based CF model. CODIGEM effectively models user-item interactions data by obtaining the intricate and non-linear patterns to generate strong collaborative signals and robust latent representations for improving the model’s generalizability and recommendation performance. Empirically, we demonstrate that CODIGEM is a very efficient generative CF model, and it outperforms several classical CF models on several real-world datasets. Moreover, we illustrate through experimental validation the settings that make CODIGEM provide the most significant recommendation performance, highlighting the importance of using the DDPM in recommendation systems.