ϵ-fair Subspace Dimensionality Reduction: An Evaluation of Fair Principal Component Analysis and Fair Column Subset Selection

Abstract

The application of artificial intelligence (AI) and machine learning (ML) is widespread presently, with many decision-makers adopting machine learning as a tool to automate the decision-making processes. As human lives are being increasingly affected by machine learning in several ways, there is a growing concern about the manner in which AI handles sensitive information to make critical decisions. Susceptibility to bias and unfairness of machine learning algorithms is evident in many recent studies, either deliberate or non-deliberate. Therefore, fairness has been taken into consideration when designing a machine learning system.

However, fairness in unsupervised learning has been largely neglected compared to supervised learning. In fact, unsupervised learning is frequently adopted as the very first step of machine learning pipelines so that bias might be unintentionally introduced therein. For example, dimensionality reduction is usually conducted to process high-dimensional data. As a result, the processed data might be biased against a protected group, even though the original is not. Currently, there are two primary schools of research into fair dimensionality reduction to combat such discriminatory behavior: one attempts to relate the dimensionality reduction with the downstream classification tasks, while the other approach, termed ϵ-fair subspace dimensionality reduction in this thesis, involves solely the dimensionality reduction task rather than any downstream tasks, concerning the reconstruction errors the projection incurs for sensitive classes.

Thus, this thesis aims to evaluate the impact of two ϵ-fair subspace dimensionality reduction methods, fair principal component analysis and fair column subset selection, on the downstream classification task by using a novel fairness measure class called ∆A-fairness and reviews recent advances in studies of fairness definitions for dimensionality reduction. The obtained results indicate that the two fair variants do not alleviate the discriminatory behavior of the vanilla variants in the downstream classification task, and the fair CSS method even performs worse on a data set comprising numerous categorical variables.