The CRISP-ML approach to handling causality and interpretability issues in machine learning

Interpretability in machine learning projects and one of its aspects - causal inference - have recently gained significant interest and focus. Due to the recent rapid appearance of frameworks, methods, algorithms and software most of which are in early stages of their development, it can be confusing for practitioners and researchers involved in a machine learning project to choose the best approach and set of techniques that would efficiently deliver valid insights while minimising the known risks of failure of data-related projects. CRISP-ML process methodology minimises this confusion by outlining a clear step-by-step process that explicitly treats of interpretability issues through every stage. The paper presents an update of CRISP-ML, which incorporates causality in a similar way and supports formalisation, design and implementation of specific instances of CRISP-ML process, subject to required levels of interpretability and causality of results. The approach is demonstrated on examples from the domains of credit risk, public health and healthcare.