An approach for interdisciplinary knowledge discovery: Link prediction between topics

Abstract

Predicting interdisciplinary links between topics can unveil potential interdisciplinary knowledge relationships and foster innovation. Considering keywords extracted from interdisciplinary research as topics, we propose a topic link prediction method based on graph neural networks. We emphasize the integration of topic semantic content features, author direct-collaboration features, and indirect-collaboration features to improve prediction performance. The interdisciplinary topic link prediction models are constructed using Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), Graph Sample and Aggregate (GraphSAGE), BERT, and Node2Vec. These models are validated by using digital humanities data as a case study. We find that the integration of semantic content, direct-collaboration, and indirect-collaboration features significantly improved the Area Under the Curve (AUC) by 20.68 % and the Average Precision (AP) by 16.52 %, compared to relying solely on the co-occurrence network. For topic reorganization, we find that the features we designed make more sense than GNN algorithms alone, and that weak relationships contribute more to topic link prediction than strong relationships. Our approach provides valuable research insights and references for scholars engaged in interdisciplinary knowledge. Notably, this is an innovative approach to interdisciplinary knowledge discovery through knowledge reorganization.