Multi-armed bandit algorithms against credit card fraud


Credit card companies are constantly on the lookout to prevent fraudulent transactions, which annually cost them millions in lost revenue. However, the large number of transactions prohibit a thorough investigation of each purchase. An expert investigator can only review a small amount of transactions per day, only a fraction of which will turn out to be fraudulent. The question then is which transactions should the expert investigate in order to minimize the lost revenue from fraud within the time budget he or she is allocated?

This problem resembles a known problem called Multi-Armed Bandit problem with Budget constraint and Variable costs (MAB-BV) [1-5]. In this thesis you will compare different MAB algorithms as part of a recommender system in order to advise human experts on the transactions they need to focus on. In addition, you will explore the effects of different types of dynamism, e.g. introducing new transaction types, changing probability of fraud, etc. Your research and insights can have significant positive impact on the revenue stream of these companies.


Supervisor: Ann NowéBernard Manderick
Contact: Kyriakos Efthymiadis (CoMo)



[1] Scott, Steven L. "Multi‐armed bandit experiments in the online service economy." Applied Stochastic Models in Business and Industry 31.1 (2015): 37-45.

[2] Ding, Wenkui, et al. "Multi-Armed Bandit with Budget Constraint and Variable Costs." AAAI. 2013.

[3] Johnson, Kris, David Simchi-Levi, and He Wang. "Online network revenue management using thompson sampling." Available at SSRN (2015).

[4] Xia, Yingce, et al. "Thompson sampling for budgeted multi-armed bandits."arXiv preprint arXiv:1505.00146 (2015).

[5] Auer, Peter, et al. "Gambling in a rigged casino: The adversarial multi-armed bandit problem." Foundations of Computer Science, 1995. Proceedings., 36th Annual Symposium on. IEEE, 1995.