Enhancing Recommendations with Articulated User Rationales

Contemporary recommendation systems primarily rely on implicit indicators of user preferences, such as clicks and viewing duration. However, these metrics often fail to explain the underlying reasons behind users' likes or dislikes. This research introduces articulated user rationales (AURs), which are user-generated textual explanations of their preferences, as a valuable source of insights for recommendations. Despite their potential, AURs present challenges in practical applications due to their sparse nature, inconsistent quality, and limited coverage. To address these issues, we propose SARA (Scaling Articulated Rationales), a comprehensive framework designed to transform sparse AURs into effective recommendation signals. SARA initiates by developing a data engine that collects and refines AURs from 240 million Kuaishou Live users, resulting in the SARA-HQ dataset, which emphasizes quality and user-centricity. Subsequently, a general-purpose MLLM is adapted into SARA-7B through extensive fine-tuning and quality enhancement processes, expanding rationale generation capabilities to encompass a broader author base. Finally, SARA-Ranker incorporates these rationales into production ranking systems, utilizing advanced interaction modeling techniques. Rigorous evaluations demonstrate that SARA-7B produces more coherent and contextually relevant rationales, while SARA-Ranker enhances user engagement and minimizes negative feedback.