The task of fraud detection often boils down to outlier detection, in which a dataset is verified to find potential anomalies in the data. Traditionally, this task was deemed a manual task, where risk experts checked all transactions manually. Even though there is a technical layer, it is purely based on a rules base that scans through each transaction, and then those shortlisted as suspicious are sent through for a manual review to make a final decision on the transaction. However, there are some major drawbacks to this system:

• Organizations need substantial fraud management budgets for manual review staff.
• Extensive training is required to train the employees working as manual review staff.
• Training the personnel to manually review transactions is time consuming and expensive.
• Even the most highly trained manual review staff carry certain biases, therefore making the whole review system inaccurate.
• Manual reviews increase the time required to fulfill a transaction. The customers might get frustrated with the long wait times required to pass a credit card transaction. This may impact the loyalty of customers.
• Manual reviews may yield false positives. A false positive not only affects the sale in the process but also lifetime value generated from the customer.

Fortunately, with the rise of machine learning (ML), artificial intelligence (AI), and deep learning, it became feasible to automate the manual credit card transaction review process to a large extent. This not only saves an intensive amount of labor but also yields better detection of credit card fraud, which otherwise is impacted due to biases that human reviewers carry.

ML-based fraud detection strategies generally can be accomplished using both supervised ML and unsupervised ML techniques.

Supervised ML models are generally used when large amounts of transaction data tagged as genuine or fraud are available. A model is trained on the labeled dataset and the resultant model is then used for classifying any new credit card transactions into one of the two possible classes.

With most organizations, the problem is that labeled data is unavailable, or very little labeled data is available. This makes supervised learning models less feasible. This is where unsupervised models come into play. They are designed to spot anomalous behavior in transactions and they do not need explicit pre-labeled data to identify the anomalous behavior. The general idea in unsupervised fraud detection is to detect behavior anomalies by identifying transactions that do not conform to the majority.

Another thing to keep in mind is that fraud events are rare, and are not as common as genuine transactions. Due to the rarity of fraud, severe class imbalance problem may be seen in datasets related to credit card fraud. In other words, one would observe that 95% or more of the data in the dataset is of genuine transactions, and less than 5% of the data belongs to fraudulent transactions. Also, even if you learn about a fraudulent transaction today, the model is likely to face an anomaly tomorrow with different features. So, the problem space of genuine transactions is well known and it is pretty much stagnant; however, the problem space for fraudulent transactions is not well known and it is not constant. Due to these reasons, it make sense to deal with the fraud detection problem with unsupervised learning rather than supervised learning.

Anomaly detection is an unsupervised learning algorithm that is also termed a one-class classification algorithm. It distinguishes between normal and anomalous observations. The key principle on which the algorithm is built is that anomalous observations do not conform to the expected pattern of other common observations in a dataset. It is called a one-class classification as it learns the pattern of genuine transactions, and anything that shows non-conformance to this pattern is termed as an anomaly, and therefore as a fraudulent transaction. The following figure is an illustration showing anomaly detection in a two-dimensional space:

A simple example of an anomaly is the identification of data points that are too far from the mean (standard deviation) in a time series. The following figure is an illustration displaying the data points that are identified as anomalies in a time series:

In this chapter, we will focus our efforts on a type of unsupervised deep learning application known as AEs.