Summary

Anomaly detection is a very active field of research because what's anomalous now may not remain anomalous forever. This poses a significant challenge to designing a good anomaly detection algorithm. Although the algorithms discussed in this chapter mostly deal with point anomalies, they can be also used to detect sequential anomalies with a little bit of feature extraction.

Sometimes, anomaly detection is treated as a classification problem, and several classification algorithms such as k-NN, SVM, and Neural Networks are deployed to identify anomalous entries. The challenge, however, is to get well-labeled data. However, some heuristics are used to assign a score called the anomaly score to each data element, and then the top few with the highest anomaly scores (sometimes above a given threshold) are determined to be anomalous.

Anomaly detection has several applications, such as finding imposters using anomaly detection on keyboard dynamics, pedestrians, and landmine detection from...