<![CDATA[Ensuring fraud-free financial management through the investigation of anomalous transactions is essential for promoting transparency and integrity in government institutions. However, existing approaches often treat anomaly detection and fraud classification as separate tasks and frequently assume that all anomalies indicate fraud. This study proposes a two-stage computational fraud detection framework that integrates Deep Isolation Forest (DIF) for anomaly detection and Extreme Gradient Boosting (XGBoost) for fraud classification, with an intermediate expert validation process. The framework was applied to a dataset of 9,166 expenditure transactions from an Indonesian government institution, consisting of 894 fraudulent and 8,272 non-fraudulent records. First, DIF identifies transactions with anomalous patterns. Subsequently, three certified internal auditors with more than ten years of experience at the Ministry of Finance review the detected anomalies to determine whether they genuinely indicate fraud. The validated results are then incorporated into an enriched dataset combining original transaction attributes and anomaly-derived features, which is used to train the XGBoost classifier. Expert judgement functions as a validation mechanism that improves label reliability and reduces the risk of misclassifying anomalies as fraud. Evaluation on a held-out test set of 1,834 transactions demonstrates strong predictive performance, achieving an accuracy of 0.9984, precision of 0.98, recall of 1.00, and an F1-score of 0.99. Although the findings are limited to a single institutional environment and require validation across multiple government agencies, the proposed framework demonstrates that integrating expert interpretation between anomaly detection and fraud classification enhances fraud identification and provides a practical, auditor-aligned approach to public-sector financial oversight.]]>
