Scientometrics Computer Science Applications
Multi context citation recommendation based on heterogeneous graph and attention mechanisms
Abstract
In recent years, against the backdrop of expanding information volume, citation recommendation has established itself as a key research direction owing to its efficiency and practical utility. Yet, issues like inadequate utilization of contextual information and suboptimal metadata fusion have notably compromised the accuracy of citation recommendation results. To address the above issues, we present a multi context citation recommendation model (MulContextRec). This model not only incorporates academic texts, historical citations, and author co-authorship information, but also infers authors’ citation intentions from the text at the section location. For different types of contexts, the model designs dedicated encoders through leveraging SciBERT, heterogeneous graph and random walk sampling—these encoders fully explore patterns in literature citations and scholars’ collaboration preferences. The model employs a dual-layer attention mechanism to achieve efficient integration, in order to avoid information redundancy or key information loss from uniform fusion, and ultimately delivers accurate support for citation recommendation with significantly improved performance. Extensive experiments on the arXiv-P, arXiv-CS, ACL, and Medline datasets demonstrate that MulContextRec outperforms all baselines across four metrics on the first three datasets, while its competitive performance on Medline further verifies its cross-domain generalization capability.