Diffusion Model Turns Sparse User Histories into Better Cold-Start Recommendations

Recommender systems face a stubborn problem: when a new user arrives with almost no interaction history, the system has very little to go on. A team of researchers has now proposed a method that addresses this user cold-start challenge by combining transferable preference patterns with diffusion-based refinement, and they report consistent gains over existing baselines across three real-world datasets.

The approach, called an Intent-Guided Diffusion Model (IGDM), is built from three components. First, an intent extraction module mines latent preference prototypes from historical user interaction sequences, capturing preference patterns that can be transferred to new users. Second, an adaptive intent fusion mechanism dynamically integrates the limited behavioral representation of a cold-start user with these prototypes, weighting them according to relevance to produce a more informative fused preference representation. Third, an intent-guided diffusion module uses that fused representation as a conditional signal to guide the diffusion-based denoising process, progressively refining the sparse and potentially noisy user representation toward a more reliable preference representation.

The key idea is that a new user’s few observed interactions should not be treated in isolation. Instead of relying solely on the limited observations available for cold-start users, the framework draws on transferable preference patterns embedded in the historical interaction data of the broader user population. The authors note that existing approaches, which mainly rely on auxiliary information, meta-learning, or sparsity simulation, still depend heavily on those limited observations and rarely exploit such transferable patterns.

The team evaluated IGDM on three real-world datasets: MovieLens-1M, Last.FM, and Yelp. According to the reported experiments, IGDM consistently outperformed multiple representative baselines in terms of Recall, NDCG, and MAP, with more pronounced improvements under highly sparse interaction settings. The authors state that these results demonstrate the effectiveness of combining transferable preference priors with diffusion-based representation refinement for user cold-start recommendation.

The work, published in Neural Processing Letters as an open access article, is presented as an accepted, peer-reviewed manuscript shared early for faster access; the version is subject to further edits and will be replaced by the final Version of Record. The research team includes Lingxuan Li, Luofei Jia, Dafei Lin, Chongmin Wang, and Yongxin Shi.

For practical recommender systems, the reported findings suggest a way to make early personalization less dependent on how much data a brand-new user has generated. By conditioning a diffusion-based denoising process on a fused preference representation that blends a user’s sparse signals with prototypes extracted from other users’ histories, the framework aims to produce a more reliable preference estimate from the very first interactions. The evaluation covers Recall, NDCG, and MAP on the three datasets named above, and the reported advantage is described as largest in highly sparse settings, which is precisely where cold-start methods are most needed. As with any benchmark study, the results reflect performance on the datasets and baselines tested, and the article is currently available as a version of record in progress.

Subject of Research: A diffusion-based recommendation framework for improving personalization for users with very limited interaction data.

Article Title: Intent-Guided Diffusion with Adaptive Intent Fusion for User Cold-Start Recommendation

Article References: Li, L., Jia, L., Lin, D., Wang, C., & Shi, Y. (2026). Intent-Guided Diffusion with Adaptive Intent Fusion for User Cold-Start Recommendation. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11894-7

Image Credits: AI Generated

DOI: 10.1007/s11063-026-11894-7

Keywords: recommender systems, cold-start problem, diffusion models, user intent, personalization, MovieLens-1M, Last.FM, Yelp, machine learning, sparse data, Intent-Guided, Diffusion