Empowering Collaborative Filtering Using Probabilistic Diffusion Models

Probabilistic diffusion models, a type of deep gen-erative models, have become one of emerging topics thanks to their state-of-the-art performance in computer vision and natural language processing (NLP) domains. Due to their denoising nature, diffusion models align well with recommender systems, where noisy user-item historical interactions are given. Diffusion models can effectively recover the original interactions from corrupted ones. Nevertheless, there are various challenges in applying diffusion models to recommender systems. To under-stand the challenges and solutions, there should be a reference for researchers and practitioners working on diffusion-based recommender systems. To this end, we first provide a comprehen-sive overview of diffusion-based collaborative filtering techniques, covering both 1) standard (non-sequential) recommendation and 2) sequential recommendation. We first explain the fundamental concepts of collaborative filtering and probabilistic diffusion models, and then summarize their applications to both standard and sequential recommendation settings.