Towards Refined Classifications Driven by SHAP Explanations

Machine Learning (ML) models are inherently approximate; as a result, the predictions of an ML model can be wrong. In applications where errors can jeopardize a company's reputation, human experts often have to manually check the alarms raised by the ML models by hand, as wrong or delayed decisions can have a significant business impact. These experts often use interpretable ML tools for the verification of predictions. However, post-prediction verification is also costly. In this paper, we hypothesize that the outputs of interpretable ML tools, such as SHAP explanations, can be exploited by machine learning techniques to improve classifier performance. By doing so, the cost of the post-prediction analysis can be reduced. To confirm our intuition, we conduct several experiments where we use SHAP explanations directly as new features. In particular, by considering nine datasets, we first compare the performance of these "SHAP features" against traditional "base features" on binary classification tasks. Then, we add a second-step classifier relying on SHAP features, with the goal of reducing false-positive and false-negative results of typical classifiers. We show that SHAP explanations used as SHAP features can help to improve classification performance, especially for false-negative reduction.