E-commerce Card Fraud Detection Using Big Data/Deep Learning Tools.

Abstract Losses due to credit card fraud are estimated at about billions of dollars per day. According to MasterCard, 118 billion of dollars in sales were declined in 2014 and 9 billion on 2018. These evolutions are partly due to the massive investment on fraud detection system (FDS) by Credit Card compagnies (Kim, et al., 2018). Designing efficient card fraud detection algorithm is the key to reduce the scourge. With their capacity to recognize pattern, Machine Learning systems could help to the task. However, some typical issue slows down the process. This thesis aims to present machine learning and deep learning models through fraud card detection systems. Several kinds of machine learning philosophies are described further (Supervised, unsupervised and reinforcement). The purpose is to find a method to increase Deep learning model precision in fraud detection. Contrarily to classical classification, detection systems must be assessed following the ranking of probability providing by the model. Several experiments have been done according to four approaches. The first one was to determine the best value on hyperparameters. In a second time, multitasks learning approach has been experimented adding a second binary feature as a second output. Then thirdly, we developed the same approach but this time with two ordinal output values. Attempting an ELM approach was the final experiment. We learned from these experiments that underfitting is probably a major break to enhance the model. However, due to cost computing some experiments could not unfortunately be tested even though they were designed. Going further is possible by testing other approaches as learning rate for hyperparameter or EasyEmsemble and SMOTE instead of under-sampling.