Prototype-Based Embedding Alignment in Multi-Scenario Multi-Task Recommendation Systems

Yuhao Yang, Yang Song, Xian Wu, Jie Zhou

Abstract: Scenario-specific divergence in user behavior presents a significant barrier to unified recommendation across multiple contexts. Traditional models often conflate shared signals with contextual nuances, resulting in degraded generalization. We propose DiscoRec, a semantic decomposition framework that re-encodes user interactions into disentangled latent spaces guided by prototype semantics. DiscoRec first decouples the feature space using a conditional encoder that isolates commonalities via soft expert routing, followed by projection into prototype-based manifolds representing both universal and context-sensitive traits. An unbalanced transport-based alignment strategy ensures faithful semantic reconstruction, treating transport plans as adaptive supervisory signals. A structural constraint is introduced to encourage diversity within the shared semantic basis. Evaluations across four benchmarks and live system deployment validate the effectiveness of DiscoRec in balancing cross-scenario generality and specificity.