Diffusion models improve diversity in recommendation rankings
When Diffusion Recommenders Stop Sampling: Preserving Generative Diversity
Information Retrieval
Summary
Many recommendation systems give just one ranked list of suggestions, but often there are several good possible choices a person might like next. The authors found that popular ways to use diffusion models for recommending items tend to produce very similar results each time instead of diverse options. They designed a new method called DARTS that separates learning the item representation, generating options, and ranking them, which helps keep the generated recommendations diverse and accurate. Testing on Amazon product reviews showed DARTS creates more varied recommendations while improving ranking quality compared to other strong methods.
What this means in practice
- •For ecommerce platform teams: Provide customers with more diverse and accurate product recommendations by integrating diffusion-based sampling without sampling collapse.
- •For movie streaming operators: Enhance next-item recommendations to offer users varied viewing options reflecting multiple plausible preferences.
Authors
Louis Castres Saint Martin, Alain Rakotomamonjy, Patrick Gallinari
Abstract
A user history can admit multiple plausible next interactions, yet conventional sequential recommenders produce a single deterministic ranking. While diffusion models offer a natural way to sample alternative continuations, we show that existing target-item diffusion recommenders often suffer from severe sampling collapse, producing nearly identical outputs across independent draws. We identify two ubiquitous design choices that promote this collapse: joint learning of the latent space and denoiser, and in-trajectory supervision via discrete losses. To address these failure modes, we introduce DARTS (Diverse and Accurate Recommendation Through Sampling), a three-stage latent diffusion framework that explicitly decouples representation learning, generative modeling, and discrete ranking by learning and freezing the latent space before diffusion training and postponing ranking supervision to the final decoding stage. We further characterize sampling diversity using complementary measures of variation before and after decoding, and use posterior-predictive aggregation to exploit this diversity at inference. Across three Amazon Reviews 2023 categories, DARTS preserves substantially greater sampling diversity while outperforming strong sequential, generative-retrieval, and diffusion baselines by 7-12% on standard ranking metrics. Aggregating multiple samples further improves ranking accuracy at a fixed slate size.