Probabilistic Residual Learning for Online Recommendations

Modern deep learning-based recommender systems often function as black boxes, making them difficult to interpret or systematically improve, especially when facing cold-start scenarios or cross-domain shifts. The authors investigate whether a modular, causal-aware framework can be added to existing models to refine their predictions by explicitly accounting for latent user groups and domain-specific confounding factors.

The authors propose Probabilistic Residual Learning (PRL), a hierarchical Bayesian framework designed to be a plug-and-play enhancement for any base recommender (e.g., DLRM, NCF, or CDL). The core mechanism involves three steps:

1. Residual Modeling: Instead of predicting ratings from scratch, PRL learns the residual (the difference between the ground-truth and the base model's prediction).
2. Probabilistic Clustering: Users are grouped into latent clusters to allow for localized, targeted residual corrections.
3. Causal Debiasing: The model uses do-calculus to marginalize over domain-level confounders (such as market or age-group effects), which helps remove spurious correlations that often lead to poor generalization in cross-domain settings.

Experiments on the XMRec and MovieLens datasets demonstrate that PRL consistently improves the performance of various base recommender models. By isolating the residual signal and adjusting for domain confounders, the framework effectively handles cold-start challenges where users have minimal interaction history. Furthermore, the model automatically discovers meaningful user clusters, providing a layer of interpretability that standard deep learning models lack.

This work provides a practical path for improving large-scale recommendation systems without needing to retrain the entire underlying architecture. By treating the existing recommender as a base component and applying a causal correction layer, developers can enhance performance and robustness in heterogeneous environments, such as global e-commerce platforms where user behavior varies significantly across different markets.