Research On Key Technologies Of Multi-Layer Network Fusion Based On Representation Learning

A network ultimately is a way to encode information about the underly-ing complex system.It has important research significance and practical values for the extraction,representation and fusion of the encoded information con-tained in the ubiquitous network data.Now people have conducted researches on complex networks and their universal properties,and revealed the rich inter-actions in complex networks.Most complex systems rarely exist and work in isolation.From social networks,infrastructure,transportation systems to cells and brains,these networks are "multi-layer networks" composed of multiple interacting networks,each of which represents a specific type or view of inter-action.Multi-layer network analysis can obtain more comprehensive and valu-able information.However,the research of multi-layer network fusion faces three challenges:First,due to privacy protection,security and technical lim-itations,it is more difficult to detect and identify the same entity in different networks;second,due to the fragmentation of graph data,the completeness of multi-views,and the heterogeneity of data from multiple sources,more com-plex features,which makes network information mining more difficult.The existing multi-layer network analysis methods still have defects in information extraction and fusion,and cannot extract deeper and more complete information well enough to solve related practical problems.In this thesis,we comprehensively utilize the relevant properties of multi-layer networks and network representation learning methods to study a com-plete set of multi-layer network fusion technologies.Starting from the three as-pects of inter-layer alignment,structure fusion,and spatial embedding fusion,this paper studies a series of key technologies in multi-layer network informa-tion extraction and fusion.The main research results are as follows:1.This thesis investigates the layer alignment problem of multi-layer net-work which is the basis of cross domain recommendation,protein function in-ference,cascading failure prediction and disease prevention and control.In or-der to identify the corresponding nodes between layers,a soft alignment model based on node ranking generation is proposed.Aiming at the shortcomings of the existing alignment algorithms based on network representation learning,this paper realizes the constraint embedding that preserves the network correlation between layers and the proximity of nodes within the layers on the basis of the constructed "1 to 1" embedding constraint,and alignment of embedded space,the graph-sequence generation model and attention mechanism are used to sort the nodes in the target network,and the correct node that should be aligned is placed at the top of the sequence.2.This thesis investigates the structural information fusion of multi-layer networks which has a wide range of application scenarios in information com-pression,In order to achieve the fusion of fragmented graph data and the com-plete reconstruction of multi-view data,this paper starts from the consistency and complementarity of structural information between layers,and aims to build a more comprehensive network structure.A nonlinear structure fusion model of multilayer network is proposed.The model preserves the structural consis-tency information by saving the highly repetitive edges in multiple networks,and preserves the complementary information by preserving the less frequent and highly similar edges.At the same time,this paper further alleviates the dependence of the construction quality of the fusion network on the embed-ding quality through high-order similarity,and then obtains a new and complete graph structure.3.This thesis investigates the spatial embedding and fusion problem of preserving the topological correlation between layers in a multi-layer network which is of great significance in the research of information diffusion and re-trieval,user anomaly detection,and cross-domain recommendation.Aiming at the problem of sampling bias in the current methodes and the imbalance of sampling edge types caused by the multi-layer structure,this thesis proposes a multi-layer network space embedding and fusion framework based on adaptive sampling and implements two multi-layer network embedding methodes with cross-layer constraints.In order to quantify the inter-layer correlation in the multi-layer network,a neighbor distribution coefficient is strictly defined,and which is used to adaptive cross-layer jump.The sampled node sequence is fed to Skip-Gram model to realize the spatial embedding and fusion of nodes in the multi-layer network.4.This thesis investigates the spatial embedding and fusion problem of preserving high-order dependencies of nodes in multi-layer attribute networks which has a broad application prospect in the completion of knowledge graphs,the construction of user portraits and recommendation system.In order to pre-serve the high-order dependencies of nodes,we take into account that there are no only dependencies between nodes in the network,but also high-order dependencies between node attributes and local topological structure.In this paper,a multi-layer network unsupervised representation learning method that preserves the high-order dependencies of nodes is proposed.We design a graph convolution-deconvolution neural network to refine the dependence of differ-ent structural semantics on node attributes in graph neural network aggregation operations.This model uses the graph Auto-encoder architecture as the frame-work of global unsupervised learning to realize the reconstruction of the original attributes of the nodes,and then obtain more accurate and fine-grained spatial embedding and fusion.