Semantic Extraction and Embedding Disentanglement for Multi-Domain Recommendations

Xiaoshuang Chen, Ben Wang, Jie Jiang, Yafei Jia, Yuhao Yang

Abstract:

Modern recommendation platforms serve personalized content across various user-facing scenarios such as main feeds, topic channels, and live video streams. These interaction environments give rise to distinct user behavior patterns, which are statistically diverse yet semantically correlated. Conventional Multi-Scenario Recommendation (MSR) approaches typically rely on a combination of shared encoders for universal patterns and separate modules for context-specific modeling. However, such designs often overlook the underlying distributional shifts between scenarios, leading to entangled representations that restrict cross-scenario generalization.
To address this, we introduce a novel framework named DistProto, which re-encodes user-item interactions into distribution-sensitive latent spaces. At its core, DistProto builds dedicated prototype manifolds that model both global and local behavior semantics. Shared features are extracted using a multi-expert gating mechanism (MMoE), while scenario-dependent encoders handle fine-grained contextual variation. These embeddings are subsequently mapped into global and scenario-specific prototype spaces, yielding disentangled representations that reflect both commonality and distinction.
To guide the alignment between feature vectors and semantic prototypes, we adopt Unbalanced Optimal Transport (UOT), which softly associates samples to prototype anchors and serves as a learning signal for semantic refinement. Additionally, to preserve diversity within the global semantic space, we impose a structural orthogonality constraint over the prototype parameters. Comprehensive empirical results from four public datasets and live deployment on a short-video platform validate the effectiveness of DistProto in improving cross-scenario personalization and robustness.