Research And Implementation Of Multiplex Network Representation Learning

Graph is an important form of data modeling,which appears in various types of real scenes.Effective graph analysis can enable researchers to gain a deeper understanding of deep information in data and benefit from various scenario applications,such as recommendation,anomaly detection,and event inference.With the development of big data,traditional graph analysis algorithms require high time and spatial complexity and cannot be effectively applied in modern big data scenarios.Therefore,algorithm research has also put forward higher demands for data scale.Graph embedding and graph neural networks provide effective and efficient solutions to address the shortcomings of graph analysis algorithms.These methods convert graph data into a low dimensional space,preserving the network’s structural and attribute information.The learned low dimensional dense representations can also be more efficiently applied in downstream applications.At present,many graph representation learning algorithms have achieved good performance on homogeneous and heterogeneous graphs,but they can only handle networks with a single node relationship and cannot effectively extend their application to multiplex networks.Compared to single view networks,nodes in multiplex networks contain various types of relationships.For example,in online shopping platforms,users and products can have relationships such as browsing,adding to carts,and purchasing.Currently,research on representation learning for multiplex networks involves applying existing graph representation learning algorithms to each view,ultimately fusing the outputs of multiple views.However,this strategy cannot effectively handle consistency and complementarity information between different views.Therefore,how to handle the complex relationships within and between views in multiplex networks is the key to research.In addition,the current application of graph representation learning in real-world scenarios is mostly limited to single view network data,such as paper classification,advisor-advisee relationship mining,and anomaly detection.How to extend and apply graph representation learning to such applications is another key issue in research.In this paper,this topic has carried out research work based on multiplex network representation learning.The main work includes the following three points:1.Aiming at the problem of mining the relationship between adviser and advisee,an attention-based multiplex network structure fusion algorithm,called AMNSF,is proposed.First,AMNSF uses GCN as an encoder to learn the node embedding that contains the local structure information of the network,and then designs the distillation module to process the consistent and complementary information respectively,and then learns the information inside and between views based on the attention mechanism.At last,we use the mining of adviser-advisee relationship as a downstream task.The experimental results on several real data sets show that the algorithm is effective in the mining of adviser-advisee relationship.2.In order to improve the robustness of the model,a self-supervised multiplex network representation learning algorithm called MNECL is proposed.Firstly,the mutual information is used to learn the node embedding containing consistent information in the multiplex network,and then another independent mutual information module is introduced to learn the complementary information related to the view.The consistency embedding and complementary embedding are fused as the output of the model.In addition,the introduction of hierarchical learning makes the learning process smoother.Finally,on multiple data sets in different fields,the performance of the algorithm on multiple downstream tasks has been significantly improved.3.Design and implement a visual graph representation learning system,which provides users with an online model modeling,model training and application platform.