A Novel Anomaly Score for Isolation Forests

Isolation Forests represent a recent variant of Random Forests, specifically designed for one-class classification problems. In the original version, this method builds a set of extremely randomized trees to describe the set of points, subsequently measuring the “anomaly” of a testing point by looking at how much deep it arrives in each tree. Even if few extensions have been recently proposed – mainly aimed at improving the training stage – in most cases the anomaly score is still kept in its original formulation, which does not completely exploit all the information contained in the trained forest. This paper is focused on improving this aspect, and proposes a new approach for the computation of the anomaly score, which exploits the different information linked to the different nodes of the trees of the forest. We investigate three dif- ferent variants of the novel anomaly score, evaluating them with twelve UCI benchmark datasets, with encouraging results.