Learning Interpretable Classification Rules Using Binarization

Yuka Yoneda, Mahito Sugiyama, Takashi Washio

Jun 17, 2017 (modified: Jun 19, 2017) ICML 2017 WHI Submission readers: everyone
  • Abstract: We propose to learn classification rules from continuous data with class labels using a twostep procedure: We first binarize data points, and second perform feature selection on the binarized data to obtain minimum necessary binary features that are required to classify given data. Each binary feature represents an interval on the original feature space, hence the user can directly interpret classification rules as intersections of intervals. Although such interpretability is indispensable in machine learning applications from biology to economics, we cannot often interpret classification rules obtained by recently emerging highly accurate machine learning methods such as deep learning. We experimentally demonstrate that our method can learn simpler classification rules than a decision tree classifier while keeping reasonable accuracy.
  • TL;DR: We learn classification rules by binarization and feature selecion on binarized data.
  • Keywords: Classification, Binarization, Feature Selection, Interpretability