A Better Random Forest Classifier: Labels Guided Mondrian Forest

A novel class of Random Forests (RFs), namely Mondrian Forests (MFs), which are an ensemble of Mondrian Trees, achieves competitive performance relatively to classical Breiman RFs. They have attractive properties like performing Bayesian inference at the tree level and being trainable online. However, they perform poorly in the presence of less or low predictive power features. Thus, we propose to extend MF by using label information during splits in order to make them more accurate and robust. We showed an increase in performance when using labels during splits on four datasets where we notice a big improvement on a dataset containing many non-predictive features which is very important as feature relevancy is unknown at first. Additionally, this extension yields equal or superior performance relatively to classical RFs.