SHAPRFs: SHapley Additive eXplanation-Based Random Forests Algorithm for Classification Problems

“Tree-based ensemble algorithms” ( \({\textit{TEAs}}\) ) are extensively employed for classification and regression problems. However, existing \({\textit{TEAs}}\) lag behind the trade-off between \({\textit{TEA}}\) interpretability and achieving cutting-edge accuracy when applied to real-time applications. This paper proposes a new \({\textit{TEA}}\) for bridging the gap between real-time applications and the trade-off between \({\textit{TEA}}\) interpretability and cutting-edge accuracy. The proposed algorithm is based on the SHapley Additive exPlanation ( \({\textit{SHAP}}\) ) value, which is one of the most advanced cooperative game solutions for quantifying each feature’s contribution to the model prediction. As a result, we call our proposed algorithm the SHapley additive explanation-based Random Forest algorithm ( \({\textit{SHAPRFs}}\) ). The proposed \({\textit{SHAPRFs}}\) compute the dominion of feature alliances using the \({\textit{SHAP}}\) value to determine the significance of each feature in the dataset. Following suit, the \({\textit{SHAPRFs}}\) divide the features into two subsets based on the importance of each feature. Then it employs the Roulette wheel selection algorithm to generate decision trees (DTs) in the forest. Five benchmark datasets are used to test the \({\textit{SHAPRFs}}\) algorithm. The experimental results show that the proposed \({\textit{SHAPRFs}}\) outperform three well-known and benchmark classifiers by a significant margin, namely, Breiman random forest ( \({\textit{BRF}}\) ), deep forest, and support vector machines ( \({\textit{SVM}}\) ).