Kuaishou Unveils OneRec AI-Powered Recommendation System

In the ever-evolving landscape of AI-driven content recommendation, Kuaishou, a leading short video platform, has introduced OneRec, a novel end-to-end generative recommendation system. This innovative system leverages an encoder-decoder architecture, enhanced by a sparse Mixture-of-Experts (MoE) technique, to significantly boost model capacity while maintaining computational efficiency. Unlike traditional cascaded recommendation systems, OneRec employs a conversational generation approach, producing contextually coherent recommendations and refining them through an iterative preference alignment module combined with Direct Preference Optimization (DPO).

What is OneRec?

OneRec represents a paradigm shift in recommendation systems. Instead of relying on a series of individual prediction steps, it frames the recommendation task as a sequence generation problem. This allows the system to consider the entire context of a user’s interaction and generate a cohesive and relevant list of recommendations.

Key Features and Functionality:

End-to-End Generative Architecture: OneRec utilizes an encoder-decoder framework to transform the recommendation problem into a sequence generation task. The encoder consolidates the user’s historical behavior sequence, while the decoder, powered by a sparse Mixture-of-Experts (MoE) architecture, progressively generates videos that the user is likely to find engaging.
Conversational Generation Method: Departing from traditional point-by-point prediction, OneRec introduces a conversational generation method. This allows the system to generate an entire recommendation list, capturing contextual information more effectively and ensuring a smoother, more relevant user experience.
Two-Stage Training Strategy: During training, OneRec quantifies multimodal representations into token sequences, feeding them into the model. A two-stage training strategy is then employed: first, the model undergoes basic item prediction task training, followed by preference alignment through DPO.
Multimodal Tokenizer: OneRec incorporates a multimodal tokenizer to process diverse types of data, including video content, user behavior, and contextual information.

Technical Principles:

The core of OneRec lies in its innovative architecture and training methodology:

Encoder-Decoder Architecture: As mentioned, the encoder compresses the user’s entire lifecycle behavior sequence into an interest vector. The decoder then uses this vector to generate a sequence of potentially interesting videos, leveraging the sparse MoE architecture. This allows the system to better capture changes in user interests and generate a coherent recommendation list.
Sparse Mixture-of-Experts (MoE): The MoE architecture allows the model to scale its capacity without a proportional increase in computational cost. This is crucial for handling the vast amount of data and complex relationships involved in personalized recommendations.
Direct Preference Optimization (DPO): DPO is a reinforcement learning technique that directly optimizes the model’s behavior based on user preferences. This allows the system to fine-tune its recommendations to better align with individual tastes.

Experimental Validation and Performance Improvement:

The effectiveness of OneRec has been validated through rigorous online A/B testing on the Kuaishou platform. The results demonstrate a significant 1.6% increase in viewing time compared to traditional methods. This improvement underscores the potential of generative recommendation systems to enhance user engagement and satisfaction.

Conclusion:

Kuaishou’s OneRec represents a significant advancement in the field of AI-powered recommendation systems. By embracing a generative, end-to-end approach, OneRec is able to deliver more relevant, contextually aware, and engaging recommendations. The successful deployment of OneRec on the Kuaishou platform highlights the potential of this technology to transform the way users discover and consume content. Future research could explore further enhancements to the model architecture, training methodologies, and integration of new data modalities. The development of OneRec signals a promising direction for the future of personalized recommendation.

References: