On Applying Meta-path for Network Embedding in Mining Heterogeneous DBLP Network

Unsupervised network embedding using neural networks garnered considerable popularity in generating network features for solving various network-based problems such as link prediction, classification, clustering, etc. As majority of the information networks are heterogeneous in nature (consist of multiple types of nodes and edges), previous approaches for heterogeneous network embedding exploit predefined meta-paths. However, a meta-path guides the model towards a specific sub-structure of the underlying heterogeneous information network, it tends to lose other inherent characteristics. Further, different meta-paths capture proximities of different semantics and may affect the performance of underlying task differently. In this paper, we systematically study the effects of different meta-paths using recently proposed network embedding methods (Metapath2vec, Node2vec, and VERSE) over DBLP bibliographic network and evaluate the performance of embeddings on two applications, namely (i) Co-authorship prediction and (ii) Author's research area classification. From various experimental observations, it is evident that embeddings exploiting different meta-paths perform differently over different tasks. It shows that meta-path based network embedding is task-specific and can not be generalized for different tasks. We further observe that selecting particular node types in heterogeneous bibliographic network yields better quality of node embeddings in comparison to considering specific meta-path.