Explainable artificial intelligence: A global fast approach

Abstract

Decisions in the financial industry place ever-increasing reliance on artificial intelligence and machine learning (AI/ML) algorithms. These decisions span entire business lines and value chains, including customer marketing, credit underwriting, financial and capital planning, algorithmic trading and automated interaction with customers, particularly chatbots and roboadvice. The most advanced algorithms, however, are complex and inherently opaque, as they require up to hundreds of inputs, which then undergo several layers of processing that are not transparent. Such complexity and opacity raise the need for explainable AI (XAI) to understand how these algorithms produce a specific output and how they work in general. Local explain capability identifies the key determinants of a specific output while global explain capability identifies the inputs that have the highest impact on the output for the algorithm as a whole. In particular, global explain capability such as the Shapley value with computational complexity of N! is prohibitively expensive with currently available approaches. This paper presents model-agnostic approaches that provide local explain capability through counterfactuals and, most importantly, global fast explanation capability.