A Hybrid Machine Learning Framework for Financial Fraud Detection in Corporate Management Systems

Abstract

Financial fraud is still a very serious problem in the operation of the enterprise, the conventional detection mode can not keep pace for changes in the fraud model, and the interpretability in the enterprise audit is lack. To improve financial anomalies detection in enterprise from a machine learning perspective, we propose an innovative hybrid machine learning approach named as EHRN-GMM, which combines a Heterogeneous Recurrent Network (HRN) and Gaussian Mixture Model (GMM). The HRN consists of GRU-LSTM fused layers with temporal embeddings and attention, which is able to learn short and long term dependencies in company sequential transaction logs. Its results are given to a GMM to estimate the distribution of valid behaviors, and detect anomalies in a probabilistic image. Furthermore, a SHAP-based interpretability layer is added which helps construct auditor-friendly explanations and enhances the transparency and trustworthiness of the model’s decisions. The proposed methodology is evaluated on synthetic ERP datasets, real-world credit card fraud data (e.g., Vesta), and simulated audit trails, where it achieves an average AUC of 0.96, outperforming competing methods such as XGBoost and CNNs. Furthermore, the model exhibits strong concept drift adaptability facilitating by GMM updates at intervals. This paper presents a scalable, interpretable, high-performing architecture for enterprise-wide fraud surveillance that helps to fill the void between automated anomaly detection and responsible corporate governance.