Generative recommendation anticipates target features for better ranking
UniSGR-V2: Stepwise Target-Side Feature Anticipation and Crossing for Generative Recommendation
Information Retrieval
Summary
Recommender systems try to suggest items you might like based on your past behavior and item features. The authors found that current generative recommendation models lack direct access to the features of the item they are predicting, which can hurt accuracy. They created a method that guesses these item features step-by-step during the recommendation process and combines them with user data to improve suggestions. Their approach also manages errors and noise to keep predictions stable and efficient, showing improvements on both offline data and live user tests.
What this means in practice
- •For ecommerce engineers: Improve product recommendation accuracy in large-scale online stores by anticipating user-item feature interactions inside the recommendation model.$Commercial implications: Enables recommender systems that generate better targeted product suggestions, potentially increasing sales and user engagement for ecommerce platforms.
- •For recommendation system developers: Build more robust generative recommendation models that reduce error propagation and manage long user histories efficiently for better ranking.
Authors
Jun Yang, Jiasi Chen, Jiawei Sun, Ziyue Guo, Jianan Yan, Tao Luo, Xiaoyi Zeng
Abstract
Semantic-ID-based generative recommendation unifies retrieval and ranking through autoregressive decoding but differs structurally from discriminative models, which can directly access rich target-item side features. The attributes are compressed into a finite codebook during quantization, while the target item remains undetermined during decoding and its features cannot be looked up. This target-side feature vacuum weakens generation at each semantic-ID level through high conditional entropy and cascading errors, and final ranking through missing target-side feature crossing. Existing methods interact with candidate representations or generate discrete attribute tokens, but none reconstructs a continuous target-side representation inside decoding to support both stepwise generation and ranking. We propose TacGR (Stepwise Target-Side Feature Anticipation and Crossing for Generative Recommendation). Before decoding each codeword, TacGR anticipates a continuous target-side feature vector and crosses it with the semantic prefix and user behavior sequence inside the decoder. The crossed representation conditions current-level generation and is directly used for ranking. Continuous-vector regression risks a learnable target collapsing with its prediction and prediction noise propagating into generation through attention. TacGR addresses these challenges with a semantic-anchoring auxiliary tower and a unidirectional visibility mask, respectively. For extremely long industrial behavior histories, it introduces gated convolutional denoising to suppress noise and adapts the pyramid memory funnel to reduce inference cost. Offline experiments on a large-scale Lazada dataset show significant gains over OneRec, UniRec, and UniSGR. Online A/B testing improves IPV, Order, and GMV by 0.79%, 5.55%, and 3.09%, respectively.