Predicting co-word links via heterogeneous graph convolutional networks.

Area of Science:

• Information Science
• Library Science

Background:

• Co-word analysis identifies research themes and networks by examining term co-occurrence.
• Machine learning, particularly link prediction in co-word networks, aids in discovering research interactions and emerging trends.
• Existing methods struggle to simultaneously learn word co-occurrence and word-document relations within co-word networks.

Purpose of the Study:

• To propose an end-to-end deep learning model for co-word network analysis.
• To jointly learn word and document embeddings from co-word networks, incorporating document-specific information.
• To improve the prediction of potential interactions between research themes and identify emerging trends.

Main Methods:

• A heterogeneous graph convolutional network (GCN) model was developed for co-word network analysis.
• The GCN model jointly learns word and document embeddings by incorporating document-specific information.
• The model is trained using binary labels indicating the presence of co-word links.

Main Results:

• The proposed GCN-based approach achieved an AUC value of [Formula: see text].
• This significantly outperformed the best traditional machine learning method, which achieved an AUC value of [Formula: see text].
• Experiments were conducted on the Web of Science dataset within Information Science and Library Science.

Conclusions:

• The heterogeneous GCN model effectively captures both word co-occurrence and word-document relationships in co-word networks.
• This approach offers a powerful tool for discovering research trends and interactions.
• The findings demonstrate the superiority of deep learning models over traditional methods in co-word network analysis.