Citation Recommendation Employing Proximity-Based Heterogeneous Network Embeddings

The number of research papers is growing exponentially on the Web and digital libraries, which makes it a cumbersome chore to determine relevant research works. To address this problem, citation recommendation (CR) models have been proposed. Nevertheless, these CR models are limited in considering the semantic relations among network objects, e.g., authors, papers, tags, venues, and topics in the heterogeneous paper’s network. Moreover, existing models do not consider the significance of proximity information between network nodes. Additionally, the current CR models face cold-start paper problems. To alleviate such problems, this work proposes a proximity-based heterogeneous network embedding (CR-PHNE) model that exploits semantic information of a network from node sequences using a probability-sensitive meta-structure-guided random walk method. Next, this information is given as input to deep neural networks to learn the latent representations of contributing nodes. Compared to its counterparts, the results produced by CR-PHNE over publicly available datasets bring 5% and 4% improvement regarding MAP and nDCG metrics, respectively. Further, the model demonstrates 7% and 3% improvements in terms of MAP and Recall@100 scores, respectively, to mitigate the cold-start paper problem compared to its counterparts.