Reproduce “Re-weighting Negative Samples for Model-Agnostic Matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiazhen Lou, Hong Wen, Fuyu Lv, Jing Zhang, Tengfei Yuan, Zhao Li
Researchers from Alibaba Group, The University of Sydney, and Zhejiang University developed UMA², an approach for recommender systems that re-weights negative samples from varying sub-spaces to mitigate sample selection bias. The method improved matching model performance by up to 0.92% in offline metrics and achieved a 3.35% Click-Through Rate gain in online A/B tests on Alibaba's platform.
Key insights
- Negative samples originating from distinct sub-spaces have varying utility and bias, indicating they should not be treated uniformly during model training.
- Applying Inverse Propensity Weighting, informed by auxiliary propensity prediction tasks, effectively mitigates Sample Selection Bias across the entire item space in recommender systems.
- A model-agnostic debiasing framework can be designed to seamlessly integrate with various existing two-tower matching architectures, broadening its applicability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.