Bias Proxy

DataFunSummit

Mar 27, 2023 · Artificial Intelligence

Model-Agnostic Self-Sampling and Bias‑Proxy Decoupling Frameworks for Debiasing Recommendation Systems

This article presents two model‑independent debiasing solutions for recommendation systems—a self‑sampling, self‑training, self‑evaluation framework (SSTE) that balances prediction accuracy and unbiasedness, and a bias‑proxy representation decoupling framework that leverages expert‑selected proxy features to remove harmful bias while preserving useful signals, with extensive offline and online evaluations in financial product recommendation scenarios.