Abstract: Forest-based methods, ensembles of decision tree classifiers, are among the top-performing classification models for many tasks. Their performance, speed, and scalability made them a perfect solution for many real-world problems. Among the best forest methods, those based on the philosophy of random forest are arguably the most successful. Their foundation is the introduction of high degrees of randomness in the induction of the decision trees of the ensemble. In this paper, we introduce a novel forest method that maximizes the potential of this philosophy. Our method constructs the forest without training the trees. Each tree is randomly constructed, and each leaf node is assigned the majority class of the instances that reach it. An even more random version assigns the class for each node using probability. Researchers have studied such purely random forests from a theoretical perspective, but their performance has consistently lagged behind that of other trees. Extensive experimentation using 400 datasets and almost all previous forest methods shows the superior performance of our approach while keeping the runtime complexity of purely random forests.
Authors: Nicolás E. García-Pedrajas, Rafael del Castillo-Gomariz, and Javier Pérez-Rodríguez
Supplementary material: