Interpretability via Random Forests

Although there is no consensus on a precise definition of interpretability, it is possible to identify several requirements: “simplicity, stability, and accuracy”, rarely all satisfied by existing interpretable methods. The structure and stability of random forests make them good candidates to improve the performance of interpretable algorithms. The first part of this chapter focuses on rule learning models, which are simple and highly predictive algorithms, but very often unstable with respect to small data perturbations. A new algorithm called SIRUS, designed as the extraction of a compact rule ensemble from a random forest, considerably improves stability over state-of-the-art competitors, while preserving simplicity and accuracy. The second part of this chapter is dedicated to post-hoc methods, in particular variable importance measures for random forests. An asymptotic analysis of Breiman’s MDA (Mean Decrease Accuracy) shows that this measure is strongly biased using a sensitivity analysis perspective. The Sobol-MDA algorithm is introduced to fix the MDA flaws, replacing permutations by projections. An extension to Shapley effects, an efficient importance measure when input variables are dependent, is then proposed with the SHAFF algorithm.