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

TL;DR

This paper investigates how different meta-paths influence network embeddings in heterogeneous DBLP networks, revealing task-dependent performance and the superiority of full-network embeddings over meta-path-based ones.

Contribution

It systematically studies the effects of various meta-paths on embedding quality across multiple methods and tasks, highlighting their task-specific nature and limitations.

Findings

Meta-path performance varies across tasks.

Embedding with all node and relation types outperforms meta-path-based embeddings.

Meta-paths are task-dependent and not universally applicable.

Abstract

In recent time, applications of network embedding in mining real-world information network have been widely reported in the literature. Majority of the information networks are heterogeneous in nature. Meta-path is one of the popularly used approaches for generating embedding in heterogeneous networks. As meta-path guides the models towards a specific sub-structure, it tends to lose some hetero- geneous characteristics inherently present in the underlying network. In this paper, we systematically study the effects of different meta-paths using different state-of-art network embedding methods (Metapath2vec, Node2vec, and VERSE) over DBLP bibliographic network and evaluate the performance of embeddings using two applications (co-authorship prediction and authors research area classification tasks). From various experimental observations, it is evident that embedding using different meta-paths perform differently over different tasks. It shows that meta- paths are task-dependent and can not be generalized for different tasks. We further observe that embedding obtained after considering all the node and relation types in bibliographic network outperforms its meta- path based counterparts.