PCA method finds hidden data biases without group labels
Identifying Representational Biases in Datasets Using PCA: A Max-Disparity Partition Framework
Machine Learning
Summary
PCA is a popular method that simplifies data but can give better results for majority groups than minority groups. The authors introduce a way to find which two groups in a dataset have the biggest difference in how well PCA represents them, without needing to know the groups beforehand. They use a special algorithm to find this split and then explain which features cause the disparity. This helps detect and address hidden biases in data automatically.
What this means in practice
- •For data science teams: Identify unnoticed representational disparities in datasets to improve fairness without requiring pre-labeled group data.
- •For machine learning engineers: Automatically discover feature-based groupings that suffer from bias under PCA, enabling targeted bias mitigation pipelines.
Tested on one dataset.
Authors
Arjun KM, Shashi Jain
Abstract
Principal Component Analysis (PCA) minimises aggregate reconstruction error, which can inadvertently represent majority subgroups with substantially higher fidelity than minority subgroups. Fairness-aware extensions of PCA correct this disparity but require group labels as input. We address the logically prior question: given only a data matrix, which binary partition of the data suffers the greatest representational disparity under a shared PCA projection? We formalise this as the max-disparity partition problem and propose a greedy local-search algorithm, grounded in the Fiduccia-Mattheyses bipartitioning framework, that discovers the disparity-maximising partition without any predefined group labels. Two benchmark algorithms, a fixed-projection sorting baseline and a simulated-annealing variant, confirm that the greedy solution is empirically near-optimal. Having identified the partition, we attribute the disparity to specific features via PCA loading scores and association rule mining, enabling a practitioner to assess whether the disadvantaged group corresponds to a human-meaningful minority. On the Predict Students' Dropout and Academic Success dataset, representational disparity is driven predominantly by institutional and programmatic proxies for socioeconomic disadvantage, with gender emerging as a secondary but consistent contributor within the disadvantaged group. The discovered partition is then passed directly to Fair PCA, completing a detect-explain-mitigate pipeline.