智通财经网·2025-09-23 11:25Core Insights - The article discusses the introduction of OneSearch by Kuaishou, an innovative end-to-end generative framework for e-commerce search, aimed at addressing the limitations of traditional cascading search architectures [1][2]. Group 1: Challenges in Traditional E-commerce Search - Traditional e-commerce platforms utilize a "recall, rough ranking, fine ranking" cascading search structure, which, while stable, faces issues such as chaotic product descriptions, relevance problems, bottlenecks in the cascading structure, and cold start challenges [1]. Group 2: OneSearch Framework Innovations - OneSearch integrates three major innovations: - Keyword Hierarchical Quantization Encoding (KHQE) module, which models product features in both vertical and horizontal dimensions, creating a rich semantic "smart ID" for each product, enhancing retrieval accuracy [2]. - Multi-perspective User Behavior Sequence Injection Strategy, which captures both recent preferences and long-term interests, allowing for a deeper understanding of user intent and improving personalized search accuracy [2]. - Preference-Aware Reward System (PARS), which combines multi-stage supervised fine-tuning with adaptive reward reinforcement learning to capture fine-grained user preference signals, enhancing ranking performance while ensuring diversity in generated results [2]. Group 3: Performance Metrics and Results - Offline experiments indicate that OneSearch significantly outperforms existing cascading systems across various metrics. Online deployment results show a 3.22% increase in order volume and a 2.4% growth in the number of buyers, marking a significant achievement in large-scale industrial applications [4]. - In a comparative analysis, OneSearch demonstrated improvements in metrics such as Click-Through Rate (CTR) and Conversion Rate (CVR), with notable enhancements in overall page satisfaction and product quality compared to traditional systems [5]. - OneSearch also excelled in cold start scenarios, effectively addressing the ranking challenges for long-tail users and newly listed products, indicating its robustness in diverse conditions [6]. Group 4: Future Directions - Kuaishou plans to continue exploring online real-time encoding solutions to bridge the gap between predefined encoding and streaming training, while also integrating more powerful reinforcement learning mechanisms to better match user preferences [6]. - The ongoing technological advancements are expected to lead to a more intelligent, precise, and user-centric e-commerce search experience, fulfilling the ideal of "one-step" search for users [6].