Independent and Sequential Ensemble Methods for Anomaly Detection

Anomaly detection problems are particularly important in various real-word contest such as fraud detection, finance, intrusion detection and cyber-security. Several methods that are presented for outlier detection work well in peculiar fields but fail if they do not meet some characteristics. We develop algorithms that can be applied in a several areas and can help in solving many real problems. In this work we focus on ensemble methods for anomaly detection in static dataset to show that combining different base learners we achieve better performance than most of the base algorithms. Before analyzing the different combinations, i.e. the ensemble strategies, we present some base algorithms being the basic components of such methods. We present and analyze different ways to combine the different base learners such as the score averaging method, the maximum score combination, the averaging ranking approach and majority voting. The ensemble methods are classified in independent and sequential. In the independent ensemble methods the base learners, that are assumed to be independent of each other, are applied to the entire dataset and the obtained scores are combined using one of the methods mentioned above. In the sequential ensemble methods the base learners are applied sequentially and every other data in the base learner is having some dependency on previous data. We report and discuss the results obtained from the implemented ensemble methods and we compare them with those obtained using basic anomaly detection algorithms. The best performances are obtained by combining different algorithms. Some of the base learners used are present in the Python libraries, others have been implemented by us.