Ensemble methods to adjust for concept drift in credit card fraud detection

Abstract

Machine Learning systems are gaining a lot of popularity due to their effectiveness in multiple domains. However, the assumption of stationary in the data, taken in research, can hinder the ability for such methods to find their way into production systems, as it is very common that real world systems and data streams see their underlying distributions change (characterized as Concept Drift), making the trained system become obsolete over time. Credit card fraud in online transactions is one such area where this drift is prevalent, which creates a major problem in a number of different industrial sectors, with millions being lost every year. This work addresses the various problems that arise when trying to fight fraud in a large scale over a long period of time. An extension to the Accuracy Weighted Ensemble is proposed, with the extension consisting of a diversity maximized ensemble that holds classifiers from previous concepts. This new ensemble is used to transfer knowledge, using weight transfer in Multi-Layer Perceptron (MLP), to inform the training of new classifiers being trained at every update stage. The proposed system is evaluated and compared to 3 other popular ensemble models that are robust to concept drift, on a fraud dataset, given by an industrial partner, as well as 4 other benchmark concept drift datasets. The result of this work shows that the proposed ensemble mechanism does provided an edge in performance compared to the other methods tested, while keeping a constant the memory footprint over time.