Generative recommendation improves video watch time with utility awareness
Utility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided Decoding
Machine Learning
Summary
Recommendation systems suggest videos or items you might like. The authors found that usual recommendation codes focus mostly on what the items are about but often miss how useful or popular they are based on user actions like shares and watch time. They created a new method that includes these usefulness signals directly in the recommendation codes and adjusts how suggestions are generated, so better and more relevant videos are less likely to be skipped. Their tests on a real platform showed this method can increase how long people watch videos.
What this means in practice
- •For video platform engineers: Improve personalized video recommendations by integrating real user engagement signals directly into recommendation generation.$Commercial implications: Enables video platforms to sell more engaging video recommendations, leading to longer user watch times and revenue increases.
- •For e-commerce recommendation teams: Enhance product suggestion quality by incorporating diverse user interaction metrics into the item encoding and decoding processes.
Authors
Yang Hu, Duo Wang, Ning Li, Kang Li, Junda She, Chengcheng Guo, Kuo Cai, Bin Liu, Youhua Liu, Xiaoshuang Chen
Abstract
Generative recommendation has been widely adopted in large-scale industrial recommendation systems, where multimodal item embeddings are quantized into offline semantic identifiers (SIDs) and items are retrieved by autoregressively generating their SIDs. However, conventional SID construction often underrepresents item utility reflected in post-exposure behaviors such as shares, long views, and video completion, making semantically similar items with different utility insufficiently distinguishable while struggling to accommodate rapidly changing utility signals. Moreover, generation likelihood may be misaligned with downstream utility, causing high-utility candidates to be prematurely pruned during beam search before post-generation reranking or training-time reward alignment can effectively compensate. To address these limitations, we propose UAGR, which aligns both SID construction and decoding with downstream multi-objective utility. Our Utility-Enriched Semantic Codes preserve the original semantic information while incorporating item-level utility without increasing SID length or vocabulary size, supporting incremental utility updates with no additional tokenization or sequence-decoding cost. Meanwhile, our Utility-Guided Decoding estimates candidate utility by combining predictions of multiple post-exposure outcomes and incorporates these estimates into beam scoring, reducing the risk of prematurely pruning high-utility paths while enabling flexible utility control without interfering with generative backbone training. Extensive offline experiments and online A/B tests under real-world traffic demonstrate the effectiveness and deployability of UAGR, which has been fully deployed on Kuaishou and increases video watch time by 0.206\% online.