Abstract
Existing methods often adjust representations adaptively only afteraggregating user behavior sequences. This coarse-grained approach tore-weighting the entire user sequence hampers the model's ability to accuratelymodel the user interest migration across different scenarios. To enhance themodel's capacity to capture user interests from historical behavior sequencesin each scenario, we develop a ranking framework named the Scenario-AdaptiveFine-Grained Personalization Network (SFPNet), which designs a kind offine-grained method for multi-scenario personalized recommendations.Specifically, SFPNet comprises a series of blocks named as Scenario-TailoringBlock, stacked sequentially. Each block initially deploys a parameterpersonalization unit to integrate scenario information at a coarse-grainedlevel by redefining fundamental features. Subsequently, we consolidatescenario-adaptively adjusted feature representations to serve as contextinformation. By employing residual connection, we incorporate this context intothe representation of each historical behavior, allowing for context-awarefine-grained customization of the behavior representations at thescenario-level, which in turn supports scenario-aware user interest modeling.