Outlier detection is used to identify exceptions, rare events, and other anomalous situations. Such anomalies may be needles in a haystack, but their consequences can nonetheless be quite dramatic; for instance, credit card fraud detection, identifying network intrusions, faults in manufacturing processes, clinical trials, voting activities, and criminal activities in e-commerce. Therefore, anomalies represent a high value when they are found and high costs if they are not. Applying machine learning to outlier detection problems can bring new insights and better detection of outlier events. Machine learning can take into account many disparate sources of data, and can find correlations that are too obscure for human analysis to identify.

Take the example of e-commerce fraud detection. With machine learning algorithms in place, the purchaser's online behavior, that is, website browsing history, becomes a part of the fraud detection algorithm, rather than simply the history of purchases made by the cardholder. This involves analyzing a variety of data sources, but it is also a far more robust approach to e-commerce fraud detection.

In this chapter, we will cover the following topics:

• Problems and challenges
• Suspicious pattern detection
• Anomalous pattern detection
• Working with unbalanced datasets
• Anomaly detection in time series