Generating complex explanations from machine learning models using class-contrastive reasoning

Objective One of the major limitations of most black-box machine learning models is the lack of explainability. In healthcare, explainability is important. Furthermore, most healthcare professionals do not have technical knowledge of machine learning. Consequently, it is necessary to translate the predictions of the machine learning model into an explainable narrative. Our research focuses on the healthcare domain. The goal of this study is to generate complex explanations from a black-box machine learning model applied to heaalthcare. Results Class-contrastive techniques can be used to generate explanations. In this method, class-contrastive counterfactual reasoning is applied to a machine learning model on tabular data (in health-care). The model predictions are explained by observing the changes in prediction by altering the inputs. This is visualized using heatmaps (class-contrastive heatmaps). This approach displays prediction results as visualizations (heatmaps). Our contribution is to extend class-contrastive analysis of black-box machine learning models to numeric features. Our work also allows machine learning scientists to visually inspect class-contrastive heatmaps and generate complex explanations for models. The resulting explanations (visual and text) are easier for non-technical people to follow. We show how machine learning scientists can extract complex explanations from machine learning models which can be interpreted by nontechnical audiences. Our work may be broadly applicable in domains where explainability is important.