Bridging Distributional Gaps in Multi-domain Recommendation via Global-Prototype Disentanglement

Gaode Chen, Xin Zhang, Meng Zhang, mu zi li, Qing Qing Wang, RuiwenXu

Abstract:

The heterogeneity of user engagement across distinct interface modalities—ranging from main feeds to real-time streaming contexts—poses a significant challenge in recommendation systems. While these environments share underlying user interests, the statistical divergence in data distributions often leads to suboptimal performance in standard Multi-Scenario Recommendation (MSR) frameworks. Current architectures, which typically rely on rigid parameter sharing combined with branch networks, struggle to adequately decouple scenario-specific nuances from common patterns. This failure to account for distributional shifts frequently results in feature interference and negative transfer.

To address these limitations, we introduce the Global-Prototype Disentanglement Network (GPD-Net). This framework establishes a novel paradigm based on distribution-adaptive latent spaces. Structurally, GPD-Net leverages a Mixture-of-Experts (MoE) backbone to extract generalized signals, complemented by standalone pathways for distinct scenario traits. A key innovation lies in projecting these signals into a prototype-guided manifold, yielding aligned representations that balance global coherence with local specificity. Furthermore, we incorporate an Unbalanced Optimal Transport (UOT) mechanism to dynamically align feature vectors with global anchors, treating the calculated transport plans as auxiliary supervision signals. To ensure robustness and prevent feature collapse, an orthogonality regularization is imposed on the prototype matrix. Validation via rigorous offline benchmarks on four datasets and live production A/B testing confirms that GPD-Net significantly outperforms state-of-the-art baselines by effectively mitigating distribution entanglement.