Dimension Reduction Forests: Local Variable Importance Using Structured Random Forests

Random forests are one of the most popular machine learning methods due to their accuracy and variable importance assessment. However, random only provide in a global sense. There is an increasing need for such assessments at local level, motivated by applications personalized medicine, policy-making, bioinformatics. We propose new nonparametric estimator that pairs flexible forest kernel with sufficient dimension reduction adapt regression function’s structure. This allows us estimate meaningful directional measure each prediction point. develop computationally efficient fitting procedure conditions recovery splitting directions. demonstrate significant gains our proposed over competing on simulated real problems. Finally, we apply method seasonal particulate matter concentration data collected Beijing, China, which yields measures. The presented here available drforest Python package. Supplementary materials this article online.