SubPSI: Detecting Hidden Bias Drift in Machine Learning Through Subgroup-Decomposed Population Stability Analysis.
This paper studies subgroup-decomposed population stability index (PSI) analysis for detecting bias drift in machine learning models, meaning shifts within subgroups that aggregate drift monitoring can miss. It tests the method with mathematical simulations, public-record comparisons and resampling experiments, and benchmarks it against Kolmogorov–Smirnov tests. An optimized implementation cut median processing time by 40.6% compared with standard PSI, and all checked comparisons gave the same results as the standard method. The paper also sets out the method’s limitations and where KS tests still do better. Submitted to NSEF Pakistan (National Science and Engineering Fair), Systems Software category.
Authors
- Hiba Chaudhry (ORCID: https://orcid.org/0000-0002-7949-8862)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24374393
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint