A Near-Optimal Algorithm for Debiasing Trained Machine Learning Models
This addresses bias mitigation in machine learning models, particularly for large-scale applications where post-processing is practical, though it is incremental as it builds on existing debiasing approaches.
The paper tackles the problem of debiasing trained machine learning models, including deep neural networks, by presenting a scalable post-processing algorithm that is proven near-optimal and outperforms previous post-processing methods while matching in-processing performance on benchmark datasets.
We present a scalable post-processing algorithm for debiasing trained models, including deep neural networks (DNNs), which we prove to be near-optimal by bounding its excess Bayes risk. We empirically validate its advantages on standard benchmark datasets across both classical algorithms as well as modern DNN architectures and demonstrate that it outperforms previous post-processing methods while performing on par with in-processing. In addition, we show that the proposed algorithm is particularly effective for models trained at scale where post-processing is a natural and practical choice.