Predicting Emerging Interdisciplinary Scientific Field Mergers via Dynamic Graph Self-Attention Networks and Semantic Text Embeddings

Abstract

Predicting previously unobserved relationships between scientific fields can be formulated as a temporal link-prediction problem. This study develops a reproducible framework that combines structural, semantic, and temporal information to rank field pairs that are likely to form new cross-field linkages. The empirical analysis is implemented against the publicly released Future Of Science (FOS) benchmark, which provides annual scientific-field co-occurrence graphs, temporal evaluation splits, semantic node representations, and public evaluation artifacts. The framework defines a positive event as the first observed appearance of a field-pair edge after the historical observation period. Pairwise predictors include common neighbors, Adamic–Adar, preferential attachment, Jaccard similarity, resource allocation, semantic cosine similarity, semantic distance, and temporal changes in node and neighborhood structure. A gradient-boosted classifier is trained chronologically, with a validation period used for model selection and probability calibration and a held-out future period reserved for final evaluation. Performance is assessed using AUROC, Average Precision (AP), Precision@50, Precision@100, and Brier score on the executed held-out test set, with uncertainty summarized across five independent seeds. The implementation supports alternative negative-sampling regimes and edge thresholds for future sensitivity analyses; those alternatives are not presented as executed empirical results here. Future candidate linkages are explicitly treated as model-generated rankings rather than demonstrated scientific mergers. The resulting system is intended as an interpretable early-warning tool for science-of-science research and human-led exploration of emerging interdisciplinary connections.