Anomaly detection using ensembles

Abstract

We show that using random forests and distance-based outlier partitioning with ensemble voting methods for supervised learning of anomaly detection provide similar accuracy results when compared to the same methods without partitioning. Further, distance-based outlier and one-class support vector machine partitioning and ensemble methods for semi-supervised learning of anomaly detection also compare favorably to the corresponding non-ensemble methods. Partitioning and ensemble methods would be required for very large datasets that need distributed computing approaches. ROC curves often show significant improvement from increased true positives in the low false positive range for ensemble methods used on several datasets. © 2011 Springer-Verlag.