Explainability of Machine Learning in Credit Referencing

Explainability is a fundamental technique to cast light on the internal behavior of AI/ML systems. This is true for all industries and also for the credit referencing industry. The explainability of AI/ML systems allows independent third parties like regulators to assess and develop qualified opinions. ML algorithms have a varying degree of intrinsic explainability. In addition, more accurate algorithms may lead to less explainable models. Several techniques can be used to generate explanations.

To address this topic using a holistic approach, ACCIS organized a dedicated workshop on April 6, 2022, titled “Explainability of Machine Learning in Credit Referencing”.

Elvira Oliva, Regulatory & Compliance, and Enrico Bagli, Data Science Manager, from CRIF, together with Giorgio Visani (University of Bologna), participated as speakers in the workshop, describing CRIF research activities and their value in the regulatory framework.

Giorgio and Enrico presented an overview of available explainability techniques for Machine Learning, focusing on CRIF research on the topic. In fact, CRIF proposed a statistical approach to evaluate the reliability and statistical robustness of explainability techniques.

Elvira focused her talk on the added value of explainability in the regulatory context and in light of the upcoming European Union AI Act, looking at the current legislation. From a regulatory perspective, technical explainability is a transparency enabler toward consumers, authorities, and counterparties. In the financial sector, it is required by specific regulatory provisions from both horizontal and industry legislation: “AI development should be focused on lawfulness and technical robustness. Considering the current draft of the proposal, and its obligations and requirements, explainability - helping to understand the rationale behind model outcomes - could increasingly become a compliance tool, aimed at fostering trustworthiness towards end-users” citing the ACCIS AI Task Force on the AI Act, which brings together data science and regulatory experts, with contributions from Elvira, Enrico and Davide Capuzzo (Senior Director Analytics Management).