A Novel Information-Entropy-Based Feature Extraction Method for Transaction Fraud Detection

Machine learning methods are widely used in transaction fraud detection. The information-entropy-based feature extraction methods have been employed to improve the machine learning methods for fraud detection. However, the following two aspects have been ignored in the previous works. 1) Fixed sliding window makes the detection model lose effective information or learn irrelevant information. 2) Only the relationship among transactions of the individuals is considered while the interactions between individuals are ignored. In this paper, an adaptive information entropy (ADAIE) feature extraction method is proposed for transaction fraud detection, including an adaptive selection of window size and considering both individual and group behavior. The validity of ADAIE is tested on a real transaction dataset, and experimental results show the improvement of the proposed method.