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 jeopardize company’s reputation, human experts often have to manually check alarms raised by hand, wrong or delayed decisions significant business impact. These use interpretable tools for verification predictions. However, post-prediction is also costly. this paper, we hypothesize that outputs tools, such SHAP explanations, exploited machine learning techniques improve classifier performance. By doing so, cost analysis reduced. To confirm our intuition, conduct several experiments explanations directly new features. particular, considering nine datasets, first compare performance these “SHAP features” against traditional “base on binary classification tasks. Then, add second-step relying features, with goal reducing false-positive and false-negative results typical classifiers. We show used features help performance, especially reduction.