Abstract:The enormous number of research papers on the Web motivated researchers to propose models that could assist users with personalized citation recommendations. Recently, Citation Recommendation (CR) models applying Network Representation Learning (NRL) techniques have revealed promising outcomes. Still, current NRL-based models are limited in terms of employing salient factors and relations between the objects of Multi-view Heterogeneous Networks (MHNs), hence, they failed to capture researchers' preferences. Besides, these models cannot exploit heterogeneity in the networks and hence suffer from the sparsity problems. To overcome these problems, we propose GCR-MHNE model, which employs a Multi-View Heterogeneous Network Embedding method to generate personalized recommendations. Specifically, it exploits semantic relations between papers based on citations, venue information, topical relevance, authors' information, and relevant labels to learn their vector representations. Moreover, the model captures the most influential features related to each semantic relation employing an attention mechanism. Compared to its counterparts, GCR-MHNE brings 6% and 7% improvements using the openly-available datasets in terms of Mean Average Precision and Normalized Discounted Cumulative Gain metrics, respectively. Furthermore, the proposed model outperforms its counterparts when the networks are sparse.

Global Citation Recommendation Employing Multi-view Heterogeneous Network Embedding