Toward Efficient Automation of Interpretable Machine Learning
Boris Kovalerchuk, Nathan Neuhaus · 2018
Developing more efficient automated methods for interpretable machine learning (ML) is an important and longterm machine-learning goal. Recent studies show that unintelligible "black" box models, such as Deep Learning Neural Networks, often outperform more interpretable "grey" or "white" box models such as Decision Trees, Bayesian networks, Logic Relational models and others. Being forced to choose between accuracy and interpretability, however, is a major obstacle in the wider adoption of ML in healthcare and other domains where decisions requires both facets. Due to human perceptual limitations in analyzing complex multidimensional relations in ML, complex ML must be "degraded" to the level of human understanding, thereby also degrading model accuracy. To address this challenge, this paper presents the Dominance Classifier and Predictor (DCP) algorithm, capable of automating the process of discovering human-understandable machine learning models that are simple and visualizable. The success of DCP is shown on the benchmark Wisconsin Breast Cancer dataset with the higher accuracy than the accuracy known for other interpretable methods on these data. Furthermore, the DCP algorithm shortens the accuracy gap between interpretable and non-interpretable models on these data. The DCP explanation includes both interpretable mathematical and visual forms. Such an approach opens a new opportunity for producing more accurate and domain-explainable ML models.