Customs fraud detection using a gradient boosting approach for joint classification and risk estimation.

Area of Science:

• Data Science
• Machine Learning
• Public Finance

Background:

• Customs revenue is vital for governments, necessitating advanced analytics for fraud detection.
• Traditional methods for detecting customs fraud are labor-intensive, costly, and rely heavily on expert judgment.
• Machine learning offers a solution to identify fraud and minimize revenue loss efficiently.

Purpose of the Study:

• To introduce and evaluate the Dual-learning XGBoost-Based Approach (DXGBA) for customs fraud detection.
• To demonstrate DXGBA's capability in simultaneously classifying fraudulent declarations and estimating associated revenue risks.
• To optimize customs inspections for maximum revenue recovery within limited resources.

Main Methods:

• Formulation of under-valued imports as a dual supervised learning task.
• Application of the DXGBA model for simultaneous classification and regression within a single boosting framework.
• Investigation of resampling strategies (SMOTE, RU) to address class imbalance.
• Evaluation using a benchmark customs dataset and comparison with baseline models.

Main Results:

• DXGBA successfully detects fraud and estimates revenue impact, ranking declarations by risk.
• The model recovers up to 87.98% of revenue while auditing only 10% of declarations.
• Enhancement pipelines using tree-based embeddings and autoencoder-based deep features further improved accuracy and revenue estimation.

Conclusions:

• DXGBA offers a high-performance solution for customs fraud detection and revenue risk assessment.
• The dual-learning framework enables efficient prioritization of inspections, maximizing revenue recovery.
• DXGBA outperforms existing methods, showcasing its potential for enhancing customs administration efficiency.