An Autoencoder Based Model for Detecting Fraudulent Credit Card Transaction

With the rapid growth in credit card based financial transactions, it has become important to identify the fraudulent ones. In this work, a two stage model is proposed to identify such fraudulent transactions. To make a fraud detection system trustworthy, both miss in fraud detection and false alarms are to minimized. Understanding and learning the complex associations among the transaction attributes is a major problem. To address this issue, at the first stage of the proposed model an autoencoder is used to transform the transaction attributes to a feature vector of lower dimension. The feature vector thus obtained is used as the input to a classifier at the second stage. Experiment is done on a benchmarked dataset. It is observed that in terms of F1-measure, proposed two stage model performs better than the systems relying on only classifier and other autoencoder based systems.