OneRec - Kuaishou's end-to-end generative recommendation system

What is OneRec?

OneRec is a new end-to-end generative recommendation system launched by Kuaishou. It adopts an encoder-decoder architecture and utilizes sparse Mixture-of-Experts (MoE) technology to improve model capacity while maintaining high computational performance. Unlike traditional cascaded recommendation systems, OneRec uses a conversational generation method to produce context-coherent recommendation results. It further improves recommendation quality through an iterative preference alignment module combined with Direct Preference Optimization (DPO). During training, OneRec quantizes multimodal representations into token sequences as input to the model and employs a two-stage training strategy: first, it trains for basic item prediction tasks, and then performs preference alignment through DPO.

OneRec's main functions

• End-to-end generative architectureOneRec employs an encoder-decoder architecture, transforming the recommendation problem into a sequence generation task. The encoder integrates the user's historical behavior sequences, while the decoder progressively generates videos that the user might be interested in through a sparse hybrid expert (MoE) architecture.
• Conversational generation methodUnlike traditional point-by-point prediction, OneRec proposes a conversational generation method that can generate the entire recommendation list and better capture contextual information.
• Experimental verification and effect improvementOneRec achieved a 1.6% increase in viewing time in online A/B testing on the Kuaishou platform, significantly outperforming traditional methods.

OneRec's technical principles

• Encoder-decoder architectureneRec employs an Encoder-Decoder architecture, transforming the recommendation problem into a sequence generation task. The encoder compresses the user's entire lifecycle behavior sequence into an interest vector, while the decoder progressively generates videos that the user might be interested in using a sparse hybrid expert (MoE) architecture. This allows for better capture of changes in user interests and the generation of a coherent recommendation list.
• Multimodal word segmenterOneRec pioneered a collaborative perception-based multimodal word segmentation solution, integrating multi-dimensional information such as video titles, tags, speech-to-text conversion, and image recognition to transform it into hierarchical semantic IDs. This enables more accurate modeling of user interests.
• Reinforcement learning preference alignmentOneRec introduces a reward-based preference alignment method, enhancing model performance through reinforcement learning. The model utilizes preference rewards, format rewards, and business rewards to construct a comprehensive reward system, optimizing the generated results. This enables the model to better perceive users' fine-grained preferences, improving recommendation accuracy.
• Performance optimizationOneRec significantly reduces the number of key operators through architectural restructuring and expands model capacity through a sparse MoE architecture, thereby significantly improving computational efficiency. Through training and inference optimizations, it achieves higher model floating-point utilization (MFU), resulting in a substantial improvement in computational efficiency.

OneRec's project address

• arXiv technical paper: https://export.arxiv.org/pdf/2502.18965

Application scenarios of OneRec

• Short video recommendationsOneRec performed exceptionally well in Kuaishou's main short video recommendation scenario, handling approximately 25% of requests (QPS). Through generative architecture and reinforcement learning preference alignment, OneRec achieved significant improvements in user engagement metrics such as dwell time, likes, follows, and comments.
• Local life servicesOneRec has also achieved significant results in Kuaishou's local life services scenario. A/B comparison experiments show that it drove a 21.01% increase in total transaction volume (GMV), a 17.89% increase in order volume, an 18.58% increase in the number of purchasing users, and a 23.02% improvement in new customer acquisition efficiency.
• Live content matchingOneRec is used for live streaming content matching. By dynamically generating recommendation sequences, it can better match users' real-time interests and improve user engagement and interaction in live streaming scenarios.
• Advertising placementIn the field of advertising, OneRec's generative architecture can dynamically generate ad recommendations based on users' interests and behaviors, improving ad accuracy and conversion rates.