Overview of Anomaly Detection techniques in Machine Learning
Akanksha Toshniwal, Kavi Mahesh, R. Jayashree
Abstract
In any dataset, events which deviate from the majority of regular patterns are called as rare events. These events can be any unusual activity, fraud, intrusion or suspicious abnormal event which may be harmful or helpful for the domain application which remains unidentified with a very large amount of data. These activities are called anomalies and are very important to detect because such undetected events might be any kind of attack in the network, sudden drops/increase in sales, the spread of disease, terrorist attacks. These anomalies can be identified using the techniques of anomaly detection (AD). There are many ways to detect anomalies like classification, nearest neighbor, clustering, statistical, spectral, information-theoretic and graph. This paper provides an overview survey of these different Anomaly Detection Techniques (ADT). Real-life data is not available as expected, so choosing the suitable AD algorithm depends on factors like input data, type of anomalies, output data, domain knowledge.