A fast classifier-based approach to credit card fraud detection

This thesis aims at addressing the problem of anomaly detection in the context of credit card fraud detection with machine learning. Specifically, the goal is to apply a new approach to two-sample testing based on classifiers recently developed for new physic searches in high-energy physics. This strategy allows one to compare batches of incoming data with a control sample of standard transactions in a statistically sound way without prior knowledge of the type of fraudulent activity. The learning algorithm at the basis of this approach is a modern implementation of kernel methods that allows for fast online training and high flexibility. This work is the first attempt to export this method to a real-world use case outside the domain of particle physics.