Sequential ensemble method for unsupervised anomaly detection

In data mining, anomaly detection aims at identifying the observations which do not conform to an expected behavior. To date, a large number of techniques for anomaly detection have been proposed and developed. Recently, researchers have paid their attention to ensemble methods to improve the accuracy of anomaly detection algorithms. Particularly, Sequential Ensemble Method (SEQ) proposed recently has shown significant improvement over other techniques. The idea of SEQ is to evaluate the scores of samples by using a second algorithm with respect to the first algorithm's output. In other words, an algorithm is firstly used to choose a set of the highest suspect abnormal samples (Dref) and then a second algorithm is applied to evaluate the final score of each data samples in the dataset with respect to only Dref. In this paper, we propose an improvement of SEQ by introducing a new way to build Dref that is based on the highest suspect normal samples instead of abnormal samples. The new algorithm is applied to a number of benchmark datasets. The experimental results show that the proposed method provided better and more stable performance compared to the previous version of SEQ and six individual algorithms.