ForestPrune: Compact Depth-Controlled Tree Ensembles

Tree ensembles are versatile supervised learning algorithms that achieve state-of-the-art performance. These models are extremely powerful but can grow to enormous sizes. As a result, tree ensembles are often post-processed to reduce memory footprint and improve interpretability. In this paper, we present ForestPrune, a novel optimization framework that can post-process tree ensembles by pruning depth layers from individual trees. We also develop a new block coordinate descent method to efficiently obtain high-quality solutions to optimization problems under this framework. The number of nodes in a decision tree increases exponentially with tree depth, so pruning deep trees can drastically improve model parsimony. ForestPrune can substantially reduce the space complexity of an ensemble for a minimal cost to performance. The framework supports various weighting schemes and contains just a single hyperparameter to tune. In our experiments, we observe that ForestPrune can reduce model size 20-fold with negligible performance loss.