An Ensemble Deep Learning Framework for Automated HS Code Classification in International Trade

Purpose - This study developed and evaluated an automated HS code classification framework that integrates CNN and transformer-based deep learning models for international trade, addressing the growing complexity in customs classification due to increased cross-border e-commerce and ambiguous product descriptions. Design/Methodology - A large-scale dataset comprising “HS Code Classification cases” was gathered from the Korea Customs Service. This dataset, which reflects borderline and challenging product categories, was preprocessed to extract product names and corresponding 10-digit HS codes. We developed and compared several deep learning models, including CNN, Bi-LSTM, and a transformer-based approach, and further constructed an ensemble model using a weighted soft- voting mechanism. The models were evaluated using top-1, top-3, and top-5 accuracy metrics to align with real-world customs workflows where multiple candidate codes are reviewed. Findings - Although top-1 accuracy is difficult to achieve for complex or novel goods, the correct HS code usually appears among the top-3 or top-5 predictions. The ensemble model consistently outperformed individual models, indicating that blending local feature extraction (CNN) with contextual understanding (transformer) is especially effective for high-granularity classifications. Originality/value - Unlike many studies that relied on smaller or narrowly focused datasets, this research used real customs data containing particularly difficult classification cases. By highlighting the practical utility of top-K accuracy and validating an ensemble strategy, it offers a concrete path toward more efficient and accurate HS code classification in real-world trade environments.