DiffCold AI model solves cold-start recommendation seesaw dilemma with diffusion

Cold-start item recommendation has long frustrated system designers: new items with no interaction history perform poorly, but boosting them typically hurts recommendations for established 'warm' items. A team of researchers from Shanghai Jiao Tong University and other institutions identify this as the 'seesaw dilemma,' rooted in a fundamental distributional disparity. Warm item embeddings live in a complex behavioral manifold shaped by rich interaction signals, while cold item embeddings are confined to a semantic manifold from auxiliary content. Previous models forced a rigid mapping between these inconsistent spaces, sacrificing precision on one side to help the other.

The team proposes DiffCold, a diffusion-based generative model that avoids this trade-off. Unlike GANs or VAEs, conditional diffusion allows the model to reconstruct warm item embeddings from content while preserving the underlying manifold structure. Two key innovations power DiffCold: a Retrieval-enhanced Aggregator that initializes generation by pulling semantically similar warm items, bypassing inefficient noise; and a Simulation-based Representation Alignment module that enforces distribution consistency between generated and real embeddings via contrastive learning.

Tested on three benchmarks, DiffCold consistently outperforms existing methods across all metrics, proving it can enhance cold-start recommendations without compromising warm item performance. The work has been accepted at ECML-PKDD 2026, a top European machine learning conference. For practitioners, this means recommender systems can now onboard new products or content more effectively without degrading the user experience for existing catalog items.