Edge-Concerned Embedding for Multiplex Heterogeneous Network

Network embedding (or graph embedding) has been researched and used widely in recent years especially in academic and e-commerce area. Most methods pay more attention to homogeneous networks with single-typed nodes or edges. However, networks in real world are more complex and larger, consisting of multiple types of nodes, edges and even node attributes. Existing algorithms treat these multiplex heterogeneous networks as homogeneous network, ignoring correlations among different node types and edge types even deep semantic information. In light of these issues, we developed a new framework to solve heterogeneous network embedding problems. We mainly focus on Attributed Multiplex Heterogeneous Network but our method can apply to both heterogeneous and homogeneous networks. We also propose an edge-concerned metapath strategy to guide random walk, providing walking guidance among different layers separated by edge type and then leverages a heterogeneous skip-gram model to compute overall node embeddings. We conduct quantitative experiments to evaluate our method on four public dataset: Amazon, Youtube, DBLP and Movielens. As demonstrated by experimental results, our method achieves statistically significant improvements over compared previous methods on link prediction tasks. We also explore the parameter sensitivity of our proposed model to figure out function fluctuation while tuning parameters.