快手OneRec揭秘

he research introduces OneRec, a novel generative recommender system designed to unify the traditional multi-stage "retrieve-and-rank" process into a single, end-to-end generative model. This unified approach, implemented with an encoder-decoder architecture and a sparse Mixture-of-Experts (MoE) structure for scalable capacity, overcomes the limitations of cascaded ranking systems. Crucially, OneRec employs a session-wise generation method, predicting a list of coherent items rather than just the next item, and incorporates an Iterative Preference Alignment (IPA) module using Direct Preference Optimization (DPO) tailored for recommendation sparsity to significantly enhance result quality. The model has been successfully deployed on the Kuaishou platform, demonstrating superior performance by achieving a substantial increase in watch-time metrics.

arxiv.org