Predicting co-word links via heterogeneous graph convolutional networks

Co-word analysis, which explores the co-occurrence of key terminology within a specific field, is a valuable tool for identifying research themes and their networks. Leveraging the booming machine learning models, link prediction in co-word networks makes it possible to discover potential interactions between research themes and reveal emerging trends. Nevertheless, few existing methods have explored end-to-end deep models, impeded by the limitations of text graph models in learning both word co-occurrence and word-document relations implicit in co-word networks simultaneously. In this work, we propose to use a heterogeneous graph convolutional network (GCN) modeling to jointly learn word embeddings and document embeddings directly from co-word networks, incorporating document-specific information. The learning model is supervised by the binary labels for the existence of co-word links. Extensive experiments have been conducted on the Web of Science dataset from Information Science and Library Science. Experimental results show that the AUC value of our GCN-based approach is [Formula: see text], whereas the AUC value of the best traditional machine learning method is [Formula: see text].