How the four-nodes motifs work in heterogeneous node representation?

Heterogeneous information networks (HIN), containing different types of entities with various kinds of interaction relations in between, provide richer information than homogeneous networks. Heterogeneous motifs are induced structural subgraph patterns with semantic in HINs. There has been many works using motifs to participate in the representation learning of HINs, but rarely to understand the respective influences of motifs. Due to the rich semantic information contained in heterogeneous motifs, the effects of different structures are inconsistent in network representation. In this paper, we introduce a case study on AMiner dataset, by extracting the heterogeneous motifs with various types of nodes and edges, especially four-node motifs, the relations between those motifs also are explored. During the study process, we first construct a set of motif instances identified by subgraph isomorphism algorithm as a weighted bipartite graph and then use another semantically related node type to extract target node features from pruned adjacency matrix. Next, a series of experiments are designed to evaluate the effect of each motif and the irrelevance of different motifs. Experimental results show that embeddings by our framework achieves excellent results compared with several state-of-the-art alternatives in node classification and clustering tasks.