A Bio-Inspired Framework for Machine Bias Interpretation

Machine learning algorithms use the past and the present to predict the future. But when given biased historical data, these algorithms can quickly become discriminatory. The area of machine learning fairness has emerged to detect and de-bias these algorithms, but has received widespread criticism for its one-size-fits-all approach, which allows certain cases of bias to slip through the cracks. In this study, we take a deeper look at the mechanisms by which machine learning algorithms develop harmful bias. We introduce a new method to interpret discriminatory systems, an Evolutionary algorithm for Feature Interaction (EFI), which we apply to several commonly used machine learning algorithms in two real-world problem instances: violent crime and median house price prediction. In the results, we discover several complex forms of bias including the encoding of race through other seemingly unrelated attributes. Ultimately we suggest that more informative interpretation tools such as EFI can be used to not only explain machine learning outcomes, but supplement and improve existing machine bias detection approaches to provide a more robust and in-depth ethical evaluation of machine learning algorithms.