RECAP: Reconstructing Embeddings with Context-Aware framework for Multi-Scenario Recommendation

Yuhao Yang, Jie Zhou, Xiaoshuang Chen

Abstract: Multi-scenario recommendation systems face the challenge of aligning user preferences across diverse interaction contexts such as feeds, search, and live content. These scenarios induce domain-specific behavior patterns, leading to distributional misalignment that conventional architectures fail to resolve. We propose RECAP, a distribution-sensitive recommendation method that captures semantic variance via disentangled prototype spaces. A mixture-of-experts framework extracts shared and scenario-specialized embeddings, which are then projected into global and local semantic spaces constructed by learnable prototypes. To encourage precise alignment, we utilize an unbalanced transport mechanism to match embeddings with prototype anchors, effectively guiding learning with soft pseudo-labels. An orthogonality penalty is introduced to enforce semantic diversity among global prototypes. RECAP demonstrates strong performance on both offline and online evaluations, particularly in mitigating inter-scenario drift.