Introduction

Data analysis often begins with an implicit assumption that all observations are valid, accurate, and trustworthy. But this is not always a reasonable assumption. Consider the case of credit card companies, who collect data consisting of records of charges to an individual's credit card. If they assumed that all charges were valid, they would open the door to thieves and fraudsters to take advantage of them. Instead, they examine their transaction datasets and look for anomalies – transactions that deviate from the general observed pattern. Since fraudulent transactions are not labeled, they have to use unsupervised learning to find these anomalies and prevent criminal activity.

There are many other situations in which anomaly detection is useful. For example, manufacturers may use anomaly detection methods to find defects in their products. Medical researchers may look for anomalies in otherwise regular heartbeat patterns to diagnose illnesses. IT security professionals try...