Study On Reduction And Representation Learning For Multiplex Networks

In recent years,more and more attention has been paid to the multiplex networks in the academic field.Compared with the simple network which can only describes the single relationship between different nodes,the multiplex networks can express multiple relationships between different nodes at the same time.In the complex system of real world,different individuals often have different connection relations.For example,there are often two kinds of connections between businessmen: business and friendships.Therefore,the multiplex networks are more suitable for describing the complex system in the real life.The mathematical representation of multiplex networks generally uses adjacency matrix or adjacency table,so the storage space required increases linearly with the network layers,and the computation for some network analysis such as robustness and random walk increases exponentially with the number of layers.Therefore,how to reduce the dimension of multiplex networks is an important research direction,and also an important basis for making the multiplex networks more suitable for describing large-scale scenes.For the dimensionality reduction of multiplex networks,this thesis mainly focuses on the following two aspects: using the similarity between the layers in multiplex networks,the layers with similar structure are gathered to reduce the number of layers;the other is using the representation learning to embed the node vector into the uniform dense space,so that the node vector can be directly used for other subsequent tasks.The main work of this thesis is summarized as follows:The aggregation of multiplex networks based on the similarity of networks: in order to simplify the study of multiplex networks,many methods for aggregating multiplex networks have been proposed,but the aggregation of multiplex networks still faces two challenges.The first one is how to measure the destruction of original multiplex networks' structure after they are aggregated and the other is how to meet different requirements for the integrity of network structures for different studies.For the first one,we design an efficient and novelty index named as similarity of multiplex networks(SMNs)for measuring the multiplex network's structural similarity between original and simplified multiplex networks.For the second one,a method for aggregating multiplex networks,named as AMNs,is proposed to compromise the simplification and maintain the core structure of multiplex networks.Several representative synthetic networks are used to evaluate the reliability of AMNs.Moreover,AMNs is applied to some real-life multiplex networks,including biology,human society,and transportation multiplex networks.The experimental results demonstrate positively that the proposed approach AMNs can simplify the multiplex networks effectively under different requirements.Synchronous feature learning for multiplex networks: the feature learning for multiplex networks faces two more challenges than single-layer networks.The first one is how to make full use of the connected information in different layers,and the other is how to embed the multiplex networks into a unified space.In this thesis,we propose a novel embedding method for multiplex networks to solve these two difficulties in learning the representation of multiplex networks mentioned above termed as Multi2 vec.The Multi2 vec preserves all the first-,second-and multi-order proximities in multiplex networks by optimizing corresponding objective functions.The network reconstruction adds valuable connections existing in other layers for nodes,aiming to improve the embedding quality,and the synchronous learning strategy provides a path to embed the multiplex networks into an unified space.Several tasks are presented to evaluate the effectiveness of Multi2 vec including visualization,link prediction and node classification.The experimental results show that the performance of Multi2 vec is more prominent than several state-of-the-art methods.Multiplex network representation learning based on graph convolution: graph convolution network is a neural network model for graph-structure data.In the complex network,nodes not only have the topological structure information of networks,but may also have their own specific attribute features.Graph convolution is the process of merging the topological structure of nodes with node's features to generate node eigenvectors.To solve the problem of node information sharing and unified space in multiplex networks embedding,a multiplex network representation learning method named Multisage is proposed.By improving the method of network reconstruction,the connections in multiplex networks can be propagated quantitatively and the dynamic convolution operation can embed the nodes in the unified space.Compared with other methods,Mutisage has better performance in visualization,link prediction and node classification.