Dealing with Explainability Requirements for Machine Learning Systems

Abstract

Explainability has recently been recognized as an increasingly important quality requirement for machine learning systems. Various methods have been proposed by machine learning researchers to explain the results of machine learning techniques. However, analyzing and operationalizing such explainability requirements is knowledge-intensive and time-consuming. This paper proposes an explainability requirements analysis framework using contextual goal models, aiming at systematically and automatically deriving appropriate explainability methods. Specifically, we comprehensively survey and analyze existing explainability methods, associating them with explainability requirements and emphasizing the context for applying them. In such a way, we can automatically operationalize explainability requirements into concrete explainability methods. We conducted a case study with ten participants to evaluate our proposal. The results illustrate the framework’s usability for satisfying the explainability requirements of machine learning systems.