Towards early detection of algorithmic bias from dataset’s bias symptoms: An empirical study

Abstract

Context:

The rise of AI software has made fairness auditing essential, particularly where biased decisions have serious impacts. This entails identifying sensitive variables and calculating fairness metrics based on predictions from a baseline model. Since model training is computationally intensive, recent research focuses on early bias assessment to detect bias before extensive training starts.

Objective:

This paper presents an empirical study to evaluate how dataset statistics, named bias symptoms, can assist in the early identification of variables that may lead to bias in the system. The aim of this study is to avoid training a machine learning model before assessing – and, in case, mitigating – its bias, thus increasing the sustainability of the development process.

Method:

We first identify a bias symptoms dataset, employing 24 datasets from diverse application domains commonly used in fairness auditing. Through extensive empirical analysis, we investigate the ability of these bias symptoms to predict variables associated with bias under three fairness definitions.

Results:

Our results demonstrate that bias symptoms are effective in supporting early predictions of bias-inducing variables under specific fairness definitions.

Conclusion:

These findings offer valuable insights for practitioners and researchers, encouraging further exploration in developing methods for proactive bias mitigation involving bias symptoms.