Research On Complex Network Reconstruction Algorithm Based On Deep Neural Networks

Many complex systems can be represented by networks,where the nodes in the network represent system variables and the edges represent the interactions between variables.The common types of networks include power grids,brain networks,biological networks,transportation networks,social networks and so on.In the era of big data,it has become increasingly important to study how to mine the interactions between nodes based on the data hidden behind the nodes.Therefore,network reconstruction has become one of the hot research directions.This thesis aims to investigate network reconstruction algorithms.Based on summarizing and analyzing some existing algorithms,the following issues are put forward as follows: inadequate handling ability of nonlinear structures,consideration only about the undirected connections while ignoring the causal relationships among nodes,and difficulties in extracting the connections among nodes from sparse samples based on many methods.To address these issues,this thesis combines the deep neural networks with network reconstruction and conducts the in-depth research by using two methods: Granger causality and graph neural networks.The main accomplishments of this study are summarized as follows:(1)To address the issues of the traditional VAR model,we first propose a Granger causality method called GRUGC based on a GRU model with a recurrent neural network structure.Subsequently,the regression prediction capabilities of the GRUGC model and the VAR-based ARIMA model are compared on the public dataset.In order to investigate the causal relationships between nodes in temporal networks,the simulation experiments are conducted,including linear VAR models,nonlinear VAR models,the Lorenz-96 model,the non-uniform embedding time-delay network models,biochemical reaction models,and the DREAM challenge datasets.Based on the evaluation metrics of AUROC and AUPR,the research results demonstrate that GRUGC exhibits the excellent performance on both artificial datasets and Dream datasets.The effectiveness and superiority of the model are verified.(2)The study of network reconstruction from the perspective of graph neural networks is carried out.Then we first propose an innovative RConv GRU model with incorporating a residual structure based on the graph convolutional GRU.Firstly,the performance of this model is compared with some common graph models on the benchmark datasets,and the results demonstrate that the residual structure enhances the performance of the graph neural network.Subsequently,the RConv GRU model is validated on the real-world traffic dataset Pe MSD4.Compared with some related models,the results demonstrate that RConv GRU outperforms some traditional models.Finally,RConv GRU is applied to the task of predicting molecular mutagenicity and the subgraph structures that may lead to molecular mutations is successfully extracted from the molecular graphs.