Abstract
Principal Component Analysis (PCA) is a fundamental tool for dimensionality reduction, yet it often preserves or amplifies biases inherent in datasets. This paper introduces a convex optimization framework for Fair PCA (FPCA) that ensures protected attributes are unrecoverable from the reduced data. By imposing mean and covariance constraints within a Semidefinite Programming (SDP) formulation, the authors demonstrate significant fairness improvements across synthetic and real-world datasets with manageable losses in explanatory power.
Introduction
Fairness in machine learning has primarily focused on supervised learning tasks. However, unsupervised methods like PCA are frequently used as preprocessing steps or for qualitative policy analysis. If PCA generates representations that encode protected attributes (e.g., race or gender), downstream tasks—even those designed to be fair—will inherit these biases. Standard fairness definitions like "equalized odds" are inapplicable here due to the lack of ground-truth labels. This work addresses this gap by defining fairness as information hiding: a dimensionality reduction is fair if a classifier cannot accurately predict protected attributes from the reduced data.
Methodology
The proposed Fair PCA (FPCA) framework extends the standard convex relaxation of PCA. The authors introduce two primary types of constraints to ensure group indistinguishability in the latent space:
- Mean Constraints: Align the centroids of different protected groups.
- Covariance Constraints: Match the "shape" or spread of the data across groups to prevent sophisticated classifiers from distinguishing them based on variance or skew.
These constraints are formulated as a Semidefinite Program (SDP), which guarantees global optimality. The framework is further extended to non-linear dimensionality reduction via Fair Kernel PCA (F-KPCA).
Experimental Results
The authors evaluated FPCA on synthetic data, 15 UCI datasets, and NHANES physical activity data.
Fairness and Utility Trade-offs
A key finding is the necessity of covariance matching. While mean matching helps, classifiers like SVMs can often still separate groups based on data orientation. FPCA with both constraints significantly reduces the predictability of protected attributes. For instance, on the Adult Income dataset, the fairness disparity metric ($\Delta$) was reduced from 0.54 to 0.07.
As shown in the figure above, there is an inherent trade-off: stricter fairness constraints lead to lower variance explained. However, substantial fairness gains are often achievable with minimal loss in utility.
Fair Clustering
In the NHANES dataset, standard PCA resulted in clusters with high age disparities (14.87% standard deviation in age composition). FPCA reduced this to 1.79%, ensuring clusters represent the population's age mix while still capturing meaningful physical activity trends.
Discussion and Limitations
While the SDP formulation provides mathematical rigor and global optimality, it is computationally more intensive than the Singular Value Decomposition (SVD) used in standard PCA. This may limit scalability for extremely high-dimensional datasets. Additionally, the model requires tuning hyperparameters ($\delta$ and $\mu$) to balance fairness and variance. Future work includes extending these convex formulations to other unsupervised tasks like fair regression and PCA deflation.
Conclusion
This work provides a robust, convex framework for ensuring fairness in unsupervised dimensionality reduction. By treating fairness as a constraint on information leakage, FPCA prevents downstream models from inheriting biases, paving the way for more equitable machine learning pipelines.
Summary updated