Variable interaction measures with random forest classifiers

Novel variable interaction measures with random forest classifiers are proposed. The proposed methods efficiently measure the change in classification performance due to non-linear interactions between variables by exploiting random permutation of out-of-bag samples in random forests. They can be readily extended to measure n-subset interactions in multi-class bagging ensembles with any base supervised classifiers. This paper experimentally compares pairwise versions of our measure in binary RF classifiers against Breiman´s Gini-based measure using three datasets, a toy dataset with known interactions and two biomedical datasets from the UCI ML repository, demonstrating the effectiveness of the proposed methods.