Explainable Artificial Intelligence in Data Science: Enhancing Transparency and Trust in Predictive Models

Abstract

In today's data science world, the use of ensembles, deep neural networks and other higher capacity algorithms that have made it their business to make decisions are now a primary tool in predictive modelling.Today's data science world is focused on making predictions using algorithms which are ensembles, deep neural networks, and other high capacity algorithms whose internal workings it is difficult for the user to probe. Explainable Artificial Intelligence (XAI) helps solve this problem by generating evidence that is comprehensible to humans that allows for understanding how the input influences the output of the model. In this paper, we focus on XAI as a practical layer to help increase the transparency, debugging, accountability and calibrated trust of models. A critical review of major approaches to XAI is completed as well as a demonstration of an empirical study, using the Wisconsin Diagnostic Breast Cancer dataset included with scikit-learn. A fixed stratified train-test split is used to train the logistic regression, random forest and gradient boosting classifiers. The results from the tests that were not performed demonstrate very high predictive ability; logistic regression achieves an accuracy of 0.986 and ROC-AUC of 0.998, while random forest and gradient boosting both achieve an accuracy of 0.958. The advantages of using interpretable evidence to accompany global permutation importance and local coefficient based contribution analysis are illustrated. In the paper, it is suggested that explanations should not be accepted as evidence of the fairness, causality or safety of a model. What are the attributes of trustworthy XAI: fidelity, stability, appropriateness for the audience, uncertainty awareness and human oversight. The study ends with a practical approach to incorporating explanation into the data-science lifecycle.