UniSGR-V2: Stepwise Target-Side Feature Anticipation and Crossing for Generative Recommendation
TacGR unifies generative recommendation and ranking by allowing a decoder to access target-side features during token generation. It uses Cascaded Sideinfo Anticipation to predict missing features at each step, crosses these predictions with semantic context, and employs specific mechanisms like Semantic anchoring and Unidirectional visibility masks to ensure stable training and prevent error propagation.