Comparing Anomaly Detection and Classification Algorithms: A Case Study in Two Domains

Abstract

Utilizing large data sets in practical scenarios usually requires identifying, annotating and classifying rare events or anomalies. Although several methods exists, there are two classes of algorithms: anomaly detection algorithms and classification algorithms. Both types of algorithms have different characteristics and in this paper, we set out to compare them on two cases. We use data from a neurointensive care unit and from microwave radio transmissions. We apply Isolation Forest and Random Forest algorithms to find events in the data that occur with a frequency of ca. 1%. The results show that classification algorithms (Random Forest) perform better and can achieve up to 100% accuracy, while the anomaly detection algorithms (Isolation Forest) can achieve only 73% at best. Based on the results, we conclude that it is better to invest in annotating data á priori and use classification algorithms, despite the lower costs of using the anomaly detection algorithms.