Multi-Perspective Explanation of Data Bias in AI: A Case Study
Tosin Adewumi · 2024
Explanations are largely lacking in some Machine Learning (ML) systems but having explanations is very helpful because they clarify events. In this case study, we use 7 bias metrics in the AI Fairness 360 (AIF360) libraryto explain bias in the German Credit Dataset (GCD) in a credit scoring scenario from multiple perspective. Some of the metrics applied are applicable in Natural Language Processing (NLP). Investigations reveal that bias exists inthe dataset for the Sensitive Attribute (SA) age. As a contribution, we show that having multiple perspectives (through multiple metrics) of bias gives a clearer assessment compared to a single one. We highlight some of the mitigation algorithms that are available for handling the bias. We publicly release our source code.