This paper presents a novel confidence-aware multi-view ensemble framework for unsupervised financial fraud detection in highly imbalanced transaction datasets. Traditional fraud detection systems rely heavily on labeled data and single-model approaches, limiting their adaptability to evolving fraud patterns and real-world constraints. To address these challenges, the proposed framework integrates multiple heterogeneous anomaly detection techniques, including Isolation Forest, Local Outlier Factor, One-Class SVM, Graph Neural Networks (GNN), and Autoencoders, to capture diverse behavioral, statistical, and relational fraud characteristics.

A key contribution of this work is a confidence-aware fusion mechanism that combines model agreement and uncertainty to produce robust anomaly scores. Additionally, a feature-space specialization strategy is employed to enhance ensemble diversity. To tackle class imbalance, Conditional Tabular GAN (CTGAN) is used to generate high-quality synthetic fraud samples, significantly improving detection performance. Furthermore, explainability is achieved using SHAP through a surrogate model, enabling interpretability in an otherwise black-box unsupervised system.

The framework is evaluated on the IEEE-CIS fraud detection dataset, demonstrating strong performance with a ROC-AUC improvement up to 0.8316 after augmentation. The proposed approach effectively balances accuracy, scalability, and interpretability, making it suitable for real-world financial cybersecurity applications.