Features Extraction and Structure Similarities Measurement of Complex Networks

Various models have been proposed to shed light on the evolution mechanisms of real-world complex networks (e.g., Facebook, Twitter, etc.) that can be expressed in terms of graph similarity. Generally, state-of-the-art research has assumed that complex networks in the real world are jointly driven by (i) multiplex features rather than a single pure mechanism, and (ii) a focus on either local or global features of complex networks. Nonetheless, the extent to which these characteristics interact to influence network evolution is not entirely clear. This study introduces an approach for calculating graph similarity based on a variety of graph features, including graph cliques, entropy, spectrum, Eigenvector centrality, cluster coefficient, and cosine similarity. Initially, each network structure was closely analyzed, and multiple features were extracted and embedded in a vector for the aim of similarity measurement. The experiments demonstrate that the proposed approach outperforms other graph similarity methods. Additionally, we find that the approach based on cosine similarity performs significantly better in terms of accurate estimations (i.e., 0.81 percent) of overall complex networks, compared to the Shortest Path Kernel (SPK) at 0.69 percent and the Weisfeiler Lehman Kernel (WLK) at 0.67 percent.