A structure-characteristic-aware network embedding model via differential evolution

Abstract

Network embedding is an efficient and effective approach for network analysis. Current network embedding approaches focus on learning representations based on the proximity similarity principle, which are effective in capturing strong assortative structural relations but fail to model complex relations in many real-world networks such as disassortative structure. To this end, we introduce a novel approach – structure-characteristic-aware network embedding (SCANE) model. SCANE learns multi-view structure characteristics and optimizes them by non-negative matrix factorization to faithfully encode diverse complex graph structures. Furthermore, to adaptively aggregate structure information according to the characteristics of a given network and mitigate unnecessary workload caused by Grid Search, SCANE exploits the differential evolution to joint constrain its network embedding process and raise efficiency. From the experimental results of SCANE compared with seven state-of-the-art related network embedding models on eleven real-world benchmark datasets, the effectiveness and superior graph representation ability of the developed SCANE approach is observed.