Community-Oriented Multi-Layer Network Embedding

At present,network has become an effective tool for complex system modeling and analysis.There are networks in the real world,such as social networks,technological networks and biological networks.At present,researchers have conducted extensive research on network analysis to better understand the nature of network.The traditional network representation method can not reveal the deep features of network well,so the study of network representation learning has become the focus of network analysis.The learning goal of network representation is to learn the low-dimensional high-density continuous vector of each network node,which can not only measure the spatial relationship between nodes in the network,but also reveal the potential relationship between deep network nodes.Network representation learning can be widely applied to many network analysis tasks.However,in real life,the real network we are exposed to is often represented in the form of multi-layer network,especially in the field of engineering application.The representation of single-layer network may lose part of the real information in the network,and it is obviously incomplete to only carry out low-dimensional node vectors on the single-layer network.In order to solve the problem of information loss in single-layer network,the research of multi-layer network has gained researchers' attention in recent years.Many complex multi-layer network systems are composed of multiple single-layer networks and cross-layer relational networks between different layers,each of which represents one of many possible types of interaction.A basic problem is how to extract community information in a multi-layer network.The current algorithms either decompose a multi-layer network into a single-layer network or extend the algorithm to a multi-layer network by using consensus clustering.However,these methods often ignore the dependencies between the layers,resulting in less than ideal precision.At the same time,the microscopic properties of the network,that is,the first-order and second-order approximations of nodes and the higher-order approximations of nodes,are often ignored in the relevant studies,resulting in unsatisfactory final results.In order to solve the problem of representation learning and association information fusion in multi-layer network,in this paper,based on the non-negative matrix decomposition method,the first-order approximation and second-order approximation information of nodes are retained in each layer of multi-layer network,and the higher-order association attribute is retained in the network.Through the inter-layer dependency relationship in the multi-layer network,we connect multiple single-layer networks in the multi-layer system to form a complete unified model.We have carried out network analysis experiments on two data sets,and compared with the currently applied methods,verified that our method is effective and can achieve better performance in effect.The proposed model can be applied to the multi-layer network analysis task in real engineering,and the learning nodes in real network are represented as vectors in low dimension to improve the network analysis effect in real task.