B0685
Title: A SAFE artificial intelligence approach
Authors: Paolo Giudici - University of Pavia (Italy)
Emanuela Raffinetti - University of Pavia (Italy) [presenting]
Abstract: The growing availability of data and computational power has allowed innovative developments in the field of Artificial Intelligence (AI). Nevertheless, the consideration of the possible adverse consequences of activities with a high societal impact has led policymakers and regulators to a degree of suspicion towards AI applications. This concern was also addressed in the recent regulation for a trustworthy AI: to be trustworthy, AI methodologies have to be SAFE. A SAFE application of AI must fulfil four key principles: it should be robust in terms of data and computations (Sustainability); it should lead to accurate predictions (Accuracy); it should not discriminate by population groups (Fairness); it should be human interpretable in terms of its drivers (Explainability). In agreement with the previous requirements, the aim is to provide a concrete response. Specifically, we combined the notion of explainability with those of accuracy and robustness through the formalization of new statistical methods based on the Shapley values and the Lorenz Zonoid tool. By means of our proposal, the most explainable variables can be detected for different groups of observations, allowing to narrow of the set of data to be analysed and consequently reducing the computational effort.
