Combining Auto Encoders and One Class Support Vectors Machine for Fraudulant Credit Card Transactions Detection
Mohamad Jeragh, Mousa AlSulaimi · 2018
Credit card fraud detection is an active field of research due to increasing fraud risk, particularly with the introduction of new credit card payment methods such as nearfield communication (NFC) enabled cards and online payments. Machine learning has shown positive results in recent years when used in credit card fraud detection; nevertheless, certain complexities exist when using machine learning in credit card fraud detection. These include the lack of fraudulent transaction data and the skewed nature of training data used in credit card fraud detection models, as well as selection of an adequate metric to measure a model's performance. This paper introduces a new unsupervised learning model based on the combination of an auto-encoder and one-class support vectors machine (OSVM), where an input is fed to an auto-encoder and a the input reconstruction error is produced and passed on to an OSVM to determine whether a transaction is fraudulent. The introduced model is compared with other models such as the individual use of OSVM, auto-encoders, and a third model based on the combination of OSVM and Auto encoder using an approach different than the one proposed in this paper. When the performance is measured using the geometric mean (GMean) and F1 score the newly proposed model shows comparable results with OSVM and improved results over the other two models.