Diffusion Models for Recommender Systems:

Recommender systems (RS) play a vital role in various online applications to alleviate the information overload problem and satisfy the users' information needs. Essentially, recommender systems are built as information filtering tools for personalized content distribution.

The advent of diffusion models in recommender systems marks a significant shift in both the capabilities and the underlying paradigm of content personalization. On the one hand, diffusion models improve the generalization performance of personalized content distribution for RS by tackling problems like inadequate collaborative signals, weak latent representations, and noisy data. On the other hand, diffusion models enable the paradigm shifting from simply distributing pre-existing content to proactively generating personalized content, where we can customize item contents—such as images, posters, and stickers—tailored to individual user preferences.

This evolution expands the boundaries of recommender systems, enabling them to not only identify and distribute relevant information but also to dynamically generate content that more effectively satisfies users' evolving needs and desires. In this tutorial, we will provide an in-depth walk-through of techniques in the frontier advancements for diffusion-model-based recommender systems from both perspectives of content distribution and generation. We will also discuss real-world industrial deployments and the future potential of diffusion models in recommender systems, offering insights into how they are reshaping the landscape of information entry towards the next-generation RS.