Multi-feature Relation And Structure Representation Learning Of Complex Networks

With the advent of the era of big data,the data magnitude of all kinds of complex networks is growing rapidly.However,the original network analysis methods are powerless in the face of the massive data,so the research and applications for complex networks have been widely concerned by the academia and industry in recent years.The research foundation and focus of this field are how to efficiently carry out network analysis tasks on complex networks,such as node classification,link prediction,community discovery,and so on.In order to carry out network analysis tasks,the key is network representation learning or network embedding,which aims to map each node in the network into a low-dimensional,distributed vector representation space.In this thesis,the following four tasks are systematically carried out to address three challenges faced by the existing works from the network representation: incorporation,computational efficiency and hypernetwork structure.(1)Ordinary network representation learning based on text enhancement.A text-enhanced network representation learning method is proposed,which introduces text features of the nodes to learn more discriminative network representations,which come from joint learning of both the network topology and text features,and include common influencing factors of both parties,while the negative sampling strategy is adopted to improve computational efficiency.The experimental results on three real-world datasets demonstrate that our proposed method outperforms other baseline methods.(2)Ordinary network representation learning based on community and text features.Inspired by inductive matrix completion algorithm,a community and text-enhanced network representation learning method is proposed,which incorporates community features and text features of the nodes into the process of network representation learning under the framework of matrix factorization.The experimental results on three real-world datasets demonstrate that our proposed method outperforms other baseline methods.(3)Hypernetwork representation learning based on transformation strategy.The hypernetwork is transformed into five types of ordinary networks based on ordinary graphs,namely,2-section graph,incidence graph,incidence graph plus 2-section graph,line graph plus incidence graph,line graph plus incidence graph plus 2-section graph,through the transformation strategy from hypergraph to ordinary graph,and then based on Deep Walk algorithm,the node vector representations for these ordinary networks corresponding to the hypernetwork are obtained.The experimental results on four hypernetwork datasets demonstrate that the node classification performance of 2-section graph is better than that of other graphs,and the link prediction performance of incidence graph plus 2-section graph is better than that of other graphs.(4)Hypernetwork representation learning based on hyperedge modeling.Three hypernetwork representation learning methods based on hyperedge modeling are proposed to effectively incorporate hyperedge information into the process of network representation learning,while the negative sampling strategy is adopted to improve computational efficiency.These three methods formulate the learning process of node vector representations including the network topology and hyperedge information as a joint optimization problem,which is solved by the stochastic gradient ascent(SGA)method.The experimental results on four hypernetwork datasets demonstrate that our proposed methods outperform other baseline methods.