Unveiling systematic biases in decisional processes: an application to discrimination discovery

Decisional processes are at the basis of several security and privacy applications. However, they are often not transparent and can be affected by human or algorithmic biases that may lead to systematically misleading or unfair outcomes. To unveil these biases, one has to identify which information was used to make the decision and to quantify to what extent such information has influenced the process outcome. Two classes of techniques are widely used to determine possible correlation between variables within decisional processes from observational data: (i) econometric techniques, in particular regression analysis, and (ii) knowledge discovery techniques, in particular association rules mining. However, these techniques, taken individually, have intrinsic drawbacks that limit their applicability. In thiswork, we propose an approach for unveiling biases in decisional processes, which leverages association rule mining for systematic hypothesis generation and regression analysis for model selection and recommendation extraction. We demonstrate the proposed approach in the context of discrimination detection, showing that not only it provides 'statistically significant' evidence of discrimination but it also allows for a more efficient operationalization of the recommendations extracted, upon which the decision maker can operate.