Fair Inference for Discrete Latent Variable Models: An Intersectional Approach

Abstract

It is now widely acknowledged that machine learning models, trained on data without due care, often exhibit discriminatory behavior. Traditional fairness research has mainly focused on supervised learning tasks, particularly classification. While fairness in unsupervised learning has received some attention, the literature has primarily addressed fair representation learning of continuous embeddings. This paper, however, takes a different approach by investigating fairness in unsupervised learning using graphical models with discrete latent variables. We develop a fair stochastic variational inference method for discrete latent variables. Our approach uses a fairness penalty on the variational distribution that reflects the principles of intersectionality, a comprehensive perspective on fairness from the fields of law, social sciences, and humanities. Intersectional fairness brings the challenge of data sparsity in minibatches, which we address via a stochastic approximation approach. We first show the utility of our method in improving equity and fairness for clustering using naïve Bayes and Gaussian mixture models on benchmark datasets. To demonstrate the generality of our approach and its potential for real-world impact, we then develop a specialized graphical model for criminal justice risk assessments, and use our fairness approach to prevent the inferences from encoding unfair societal biases.