Interpretable Machine Learning with Boosting by Boolean Algorithm
Department or Administrative Unit
Much research has been conducted for developing efficient, automated, interpretable, and highly accurate Machine Learning algorithms. However, the primary challenge remains - developing algorithms successful in all of these aspects simultaneously. For applications, where human lives are at stake, having to choose between accuracy and interpretability is not acceptable. To address this challenge, previously the Dominance Classifier and Predictor (DCP) algorithm, capable of discovering of the human-understandable models in simple and visualizable terms was proposed. On the benchmark Wisconsin Breast Cancer (WBC) dataset, it achieved greater accuracy than other interpretable algorithms, reducing the gap between the prediction accuracy of interpretable and non-interpretable algorithms on these data. This paper proposes a new interpretable algorithm RPPR, to bridge this gap more by boosting DCP via discovering properties of misclassified cases. Experiments with RPPR on WBC and two other benchmark datasets using 10-fold cross validation, achieved accuracies greater than 99%. Thus, the accuracy of non-interpretable algorithms is reachable without sacrificing interpretability.
Neuhaus, N., & Kovalerchuk, B. (2019). Interpretable Machine Learning with Boosting by Boolean Algorithm. 2019 Joint 8th International Conference on Informatics, Electronics & Vision (ICIEV) and 2019 3rd International Conference on Imaging, Vision & Pattern Recognition (IcIVPR). https://doi.org/10.1109/iciev.2019.8858511
2019 Joint 8th International Conference on Informatics, Electronics & Vision (ICIEV) and 2019 3rd International Conference on Imaging, Vision & Pattern Recognition (icIVPR)
Copyright © 2019, IEEE