Machine Learning For Detecting Nuclear-related Strategic Trade: An Introduction To Supervised Classification And Nlp-assisted Matching For Transaction Identification

Detecting the transfer of nuclear-related strategic goods within the international trading system is a long-standing challenge. The small volume, difficulty correlating Harmonized System codes strategic trade control lists, and the dual-use nature of many of the commodities combine to make this a tremendous challenge. Advances in trade data collection and computing capabilities now provide the opportunity to apply machine learning to this issue. This paper provides an introduction to two techniques to improve the identification of nuclear-related strategic trade transactions: supervised classification and natural language processing (NLP) with fuzzy matching. The supervised classification approach uses data resampling and classification algorithms to model common patterns and characteristics that separate transactions involving nuclear-related strategic goods from broader international trade flows. This approach creates models using historical transaction-level and export licensing data in order to predict whether new transactions are likely to contain a nuclear-related strategic good. The second technique uses NLP to operationalize strategic trade control lists and other relevant descriptions of nuclear-related commodities. After pre-processing the text, this approach uses fuzzy matching to identify nuclear-related trade based on descriptions provided in shipping documentation. Both of these approaches can be used by state authorities to improve strategic trade control enforcement and outreach for nuclear-related commodities.