What is OneRec
OneRec is a new end-to-end generative recommendation system launched by Kuaishou. It adopts an encoder-decoder architecture and enhances model capacity through sparse Mixture-of-Experts (MoE) technology while maintaining efficient computational performance. Unlike traditional cascaded recommendation systems, OneRec uses a conversational generation method that can generate contextually coherent recommendation results, and 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 model input and adopts a two-stage training strategy: first training on the basic item prediction task, then performing preference alignment through DPO.
Main Features of OneRec
-
End-to-End Generative Architecture: OneRec adopts an encoder-decoder architecture, transforming the recommendation problem into a sequence generation task. The encoder integrates the user's historical behavior sequence, and the decoder gradually generates videos that the user may be interested in through a sparse Mixture-of-Experts (MoE) architecture.
-
Conversational Generation Method: Unlike traditional pointwise prediction, OneRec proposes a conversational generation method that can generate the entire recommendation list and better capture contextual information.
-
Experimental Validation and Performance Improvement: OneRec achieved a 1.6% increase in watch time in online A/B tests on the Kuaishou platform, significantly outperforming traditional methods.
Technical Principles of OneRec
-
Encoder-Decoder Architecture: neRec adopts an Encoder-Decoder architecture, transforming the recommendation problem into a sequence generation task. The encoder is responsible for compressing the user's full lifecycle behavior sequence into an interest vector, while the decoder gradually generates videos that the user may be interested in through a sparse Mixture-of-Experts (MoE) architecture. It can better capture changes in user interests and generate coherent recommendation lists.
-
Multimodal Tokenizer: OneRec pioneered a collaborative-aware multimodal tokenization scheme that integrates multidimensional information such as video titles, tags, speech-to-text, and image recognition, converting it into hierarchical semantic IDs. It can model user interests more accurately.
-
Reinforcement Learning Preference Alignment: OneRec introduces a reward mechanism-based preference alignment method, enhancing model performance through reinforcement learning. The model uses preference rewards, format rewards, and business rewards to build a comprehensive reward system and optimize generated results. This enables the model to better perceive users' fine-grained preferences and improve recommendation accuracy.
-
Performance Optimization: Through architectural restructuring, OneRec greatly compresses the number of key operators, expands model capacity through a sparse MoE architecture, and significantly improves computational efficiency. Through training and inference optimization, it achieves higher Model FLOPs Utilization (MFU), greatly improving computing efficiency.
Project Address of OneRec
- arXiv Technical Paper：https://export.arxiv.org/pdf/2502.18965
Application Scenarios of OneRec
- Short Video Recommendation: OneRec performs outstandingly in Kuaishou's main short video recommendation scenario, handling about 25% of requests (QPS). Through its generative architecture and reinforcement learning preference alignment, OneRec has achieved significant improvements in interaction metrics such as user dwell time, likes, follows, and comments.
- Local Life Services: OneRec 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 gross merchandise 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 matching: OneRec is applied to live content matching. By dynamically generating recommendation sequences, it can better match users' real-time interests and improve user engagement and interaction rates in live streaming scenarios.
- Ad delivery: In the field of ad delivery, OneRec's generative architecture can dynamically generate ad recommendations based on users' interests and behavior, improving ad precision and conversion rates.