Algorithmic Bias and Fairness in Case-Based Reasoning

Abstract

Algorithmic bias due to underestimation refers to situations where an algorithm under-predicts desirable outcomes for a protected minority. In this paper we show how this can be addressed in a case-based reasoning (CBR) context by a metric learning strategy that explicitly considers bias/fairness. Since one of the advantages CBR has over alternative machine learning approaches is interpretability, it is interesting to see how much this metric learning distorts the case-retrieval process. We find that bias is addressed with a minimum impact on case-based predictions - little more than the predictions that need to be changed are changed. However, the effect on explanation is more significant as the case-retrieval order is impacted.