Nuclear fuel cycle-related R&D classification for implementing the IAEA's additional protocol

Abstract

Novel methodologies for classifying scientific articles related to the nuclear fuel cycle have been developed using machine learning to discover declarable activities under the Additional Protocol of the International Atomic Energy Agency. In this study, the relationships between articles and their lists of references or authors were analyzed using a network to examine the resultant features. By comparing the original network and a randomly rewired network based on the original data, we show that article topics and lists of references or authors form clusters in a projected bipartite network, indicating that lists of references or authors can be employed as independent variables for classification. The topics of scientific articles were classified using the lists of article authors, lists of references, and abstract word counts. Notably, decision-tree classifiers and logistic regression exhibit high F1_score and recall. Furthermore, to improve classifier performance, ensemble classifiers were applied based on the abovementioned single classifiers. The combined classifiers with logistic regression based on author lists as an independent variable showed a particularly high recall value when the subject of an article was distinguished. This classification method could contribute to a better understanding for determining and monitoring nuclear fuel cycle-related R&D to achieve safeguard objectives.