Research on Enterprise Fraud Detection Based on Transaction Flow ETL and Deep Anomaly Detection

Abstract

Enterprise transaction fraud identification is a crucial task for risk control and fund security management in a digital business environment. Addressing the challenges of complex transaction data sources, diverse field types, concealed anomaly patterns, and uneven distribution of fraud samples, this paper proposes an enterprise fraud identification method based on transaction flow ETL and deep anomaly detection. This method first constructs a data processing workflow for real e-commerce transaction scenarios, handling missing values, encoding categorical features, normalizing numerical values, and integrating transaction attributes in the original transaction records to form a unified and stable modeling input from multi-dimensional transaction information. Based on this, a deep anomaly detection model is used to learn the latent representation of transaction behavior. The encoder extracts nonlinear correlations from transaction features, and a fraud risk scoring mechanism is constructed by combining reconstruction bias and latent space deviation, thereby achieving automatic identification of abnormal transactions. This paper uses the IEEE-CIS Fraud Detection dataset for research. This dataset includes transaction amount, transaction time, identity information, device information, and anonymous business features, reflecting the complex risk characteristics of real e-commerce transactions. Comparative experimental results show that the proposed method exhibits superior overall performance in terms of accuracy, precision, recall, and F1 score, indicating that the combination of transaction flow ETL and deep anomaly detection can improve the stability and effectiveness of enterprise fraud identification. This research provides a feasible data-driven approach for enterprise intelligent risk control, transaction security auditing, and abnormal transaction early warning.