SIRUS: Stable and Interpretable RUle Set for classification

State-of-the-art learning algorithms, such as random forests or neural networks, are often qualified “black-boxes” because of the high number and complexity operations involved in their prediction mechanism. This lack interpretability is a strong limitation for applications involving critical decisions, typically analysis production processes manufacturing industry. In contexts, models have to be interpretable, i.e., simple, stable, predictive. To address this issue, we design SIRUS (Stable Interpretable RUle Set), new classification algorithm based on forests, which takes form short list rules. While simple usually unstable with respect data perturbation, achieves remarkable stability improvement over cutting-edge methods. Furthermore, inherits predictive accuracy close combined simplicity decision trees. These properties assessed both from theoretical empirical point view, through extensive numerical experiments our $\mathtt{R/C}\mathtt{++}$ software implementation $\mathtt{sirus}$ available $\mathtt{CRAN}$.