A Revised Isolation Forest procedure for Anomaly Detection with High Number of Data Points

Abstract

Anomaly Detection defines the process of identifying unusual data characterised by a different behaviour with respect to the rest of the dataset. It represents a key element for Industry 4.0: Anomaly Detection allows to automatically monitor and identify anomalous events in collected data with the purpose of preventing the occurrence of failures and malfunctions and of taking quick actions if they occur. Over the years, the amount of data to be analysed has grown enormously, posing a challenge for classical anomaly detectors originally created for small sets of data. This paper proposes a reinterpretation of the classical Isolation Forest procedure mainly suitable to tackle dataset characterised by high number of data points. We consider an alternative way to calculate the split value so to avoid to split in high density areas, increasing the isolation power of the method. The performance of the proposed approach is tested on several real dataset and some promising results emerge.