Disentangled User Modeling for Multi-Scenario Recommendation via Shared Behavior Embedding

Fei-Fei Xue, Liang Chen, Rui Zhang

Abstract:

Contemporary mobile applications are increasingly required to deliver personalized recommendations across diverse interfaces, such as home feeds, local services, and live streaming contexts (e.g., on TikTok). These distinct contexts engender unique user behavioral patterns, leading to highly heterogeneous data distributions. To manage these variations and boost system performance, existing Multi-Scenario Recommendation (MSR) frameworks typically rely on shared parameter mechanisms to facilitate knowledge transfer. However, direct parameter sharing often proves insufficient for achieving precise feature alignment and knowledge propagation when facing severe distributional discrepancies between scenarios.
To address this limitation, we present a novel approach termed Prototypical Knowledge Transfer (PKT), which integrates Optimal Transport logic into the MSR landscape. Technically, our framework adopts the Multi-gate Mixture of Experts (MMoE) architecture as a foundational shared feature extractor. We uniquely introduce a prototype layer that functions as an intermediary distribution, creating a semantic bridge to harmonize different scenario distributions. Furthermore, we apply Optimal Transport to optimize the efficiency of knowledge transfer from scenario-specific distributions to these prototypes. To ensure the granularity of scenario-specific representations, we deploy independent expert networks for each scenario, augmented by the LHUC architecture. The proposed method's efficacy is validated through extensive offline experiments on two datasets