Heterogeneous graphlets-guided network embedding via eulerian-trail-based representation

Abstract

Heterogeneous graph neural networks (GNNs) have excelled at learning to represent heterogeneous information networks (HINs). However, there are the following limitations in heterogeneous information networks representations: (1) existing models rely on pre-defined substructure sets, such as metapaths, in which the number of nodes is primarily determined based on experience; (2) except for a few simple pairwise relations, most models fail to capture multifaceted graph structures and commonly result in semantic redundancy. To tackle the aforementioned problems, in this study, we propose a Heterogeneous Graphlets-guided Network Embedding framework via Eulerian-trail-based Representation (HeGEER) for capturing complex structural characteristics automatically and efficiently. HeGEER first explores frequent heterogeneous graphlets through a random-algorithm-based method without any artificial pre-definition. Then, HeGEER conducts a Jaccard-coefficients-based sorting to reduce the semantical redundancy of derived heterogeneous graphlets, and transforms these heterogeneous graphlets into corresponding Eulerian-trails to preserve the entire relation information. The node embeddings are then generated by aggregating representations of intro- and inter- heterogeneous graphlets using a heterogeneous graph dual-attention mechanism. On various benchmark datasets, comprehensive experiments are undertaken on tasks including node classification and link prediction. The experimental results are thoroughly analyzed to illustrate HeGEER’s improved performance compared to several typical models.