Density-based isolation forest
Nathan Levêque, Emmanuel Frénod, Jérôme Lacaille, Audrey Poterie, François Septier, Mourad Yahia-Bacha · 2025
This work addresses the problem of outlier detection and studies isolation forest algorithms, originally introduced by Liu et al. (2008). We propose a new algorithm, which we call the density-based isolation forest (dbiForest). A dbiForest consists of a large number of density-based isolation trees (dbiTrees). Each dbiTree aims to identify outliers through recursive binary splits based on multivariate density estimates. This original approach integrates density estimation into the structure of the isolation tree. Based on the assumption that outliers are both rare and different from normal values, this algorithm identifies outliers as points that lie in regions of low density. Through several numerical experiments, we empirically show that this new method can overcome some of the limitations of existing isolation forest algorithms.