Self-Generative Feature Evolution Models For Improved Predictive Accuracy In Financial Fraud Detection Applications

Abstract

Anti-fraud detection models depend on a predefined set of engineered or learned features, however, since the fraudulent behaviour is by nature adversarial, it will try to find ways to bypass the detection border line and hence render the feature set inefficient very quickly. In this paper we present a self-generative feature evolution approach where an autoencoder first learns a small dimensional latent representation of the transaction data after which a conditional generative adversarial network is used to generate evolving fraud pattern features from a dynamically updated feedback buffer of known fraud cases and combine them with actual features for classification using a gradient-boosted ensemble model. Once drift or a new fraud pattern type is detected, the generator and autoencoder can be automatically re-trained using the new feedback buffer and the feature space can hence evolve instead of staying static after training. The proposed algorithm has been tested using a statistical replica of the ULB credit card fraud benchmark data set over four different stages. When compared with static feature engineering baseline as well as the autoencoder SMOTE baseline which does not evolve its feature space, the suggested SGFE framework maintains a fraud recall of over 86 percent in all four phases while the recall of the static feature engineering baseline falls to 58.1 percent from 87.4 percent owing to the emergence of new typologies. In general, SGFE outperforms both baseline methods in terms of precision, recall, F1 score, area under the precision-recall curve, and Matthew’s correlation coefficient.