Research Of Graph Neural Networks For Networks' Uncertainty

The rapid development of networks provides a new perspective for people to study social life.Modeling biological,social,and transportation systems as net-works can recognize patterns more efficiently through network embedding and analysis than traditional methods.However,with the continuous development of the relational data set based on network,its data acquisition and analysis are full of uncertainties,and the network analysis techniques are also challenging in terms of calculations and concepts.Especially for many techniques such as graph neural network,the features and network's structure must be known be-fore training.But in real-world tasks,these are not always given,which leads to many uncertainties,including the uncertainty of nodes' features,nodes' relation-ship and network's structure.These uncertainties not only lead to insufficient accuracy,but also result in low robustness and generalization of models based on these networks.Thus,this thesis combines the typical application scenarios to carry out innovative research and exploration on key technologies for solving network's uncertainties.The specific work and contributions are as follows:1.This thesis studies the influence relationship between different features and nodes' labels.It uses Bayes' theorem to generate the probability distri-bution between nodes' features and nodes' labels,and then predicts the unknown labels by features.It quantitatively evaluates the model with social network data,and intuitively analyzes the probability distribution between nodes'features and nodes' labels to judge the model's solution to the uncertainty of nodes' features.The experiments affirm the analysis and solution ideas for the uncertainty of nodes' features.2.This thesis studies the influence of different neighbors on the node.It first proposes a special sampling scheme to solve the problem of neighboring sampling expanding noise and weakening important edges in large-scale networks.Furthermore,it proposes a special graph neural network,which can assign different weights adaptively to different neighbors during neigh-borhood aggregation,and updates nodes' features iteratively.It evaluates the model quantitatively with citation network data.The results show that the uncertainty of nodes' relationship will affect model's effectiveness and generalization significantly.The model can solve the uncertainty of nodes'relationship and avoid overfitting effectively.3.This thesis proposes a Bayesian graph neural network to reconstruct the network by using mixed-membership stochastic block model and then use the maximum a posteriori approximate to simplify it.Meanwhile,it assigns different weights adaptively to different neighbors during the training pro-cess.Thus,it can solve the problem of the uncertainty of network's struc-ture and the uncertainty of nodes' relationship at the same time.Taking the traffic network as an example,it evaluates the model quantitatively in the uncertainty of spatio-temporal network's structure and complex depen-dence relationships.Extensive experimental results affirm the advancement,effectiveness and high robustness of the model.4.In view of the problem of mixed uncertainties in real-world scene,this the-sis takes the air spatio-temporal network as an example to study the solu-tions.It models the spatio-temporal network as a sequence of heterogeneous graphs.Then it uses the heterogeneous graph-level attention to solve the uncertainty of spatial dependence relationship and uses the sequence-level attention to solve the uncertainty of temporal dependence relationship.At the same time,it uses a convolutional subnet to control the importance of different head extraction features in the multi-head attention to solve the uncertainty of nodes' features.Then starting with the case of multi-ple uncertainties-flight delays,it uses extensive experiments to prove the effectiveness of the model,and affirm the effectiveness of solving network uncertainties in improving the model.In summary,this thesis conducts in-depth analysis and research on the key technologies and methods to eliminate network uncertainty.The proposed cor-relative graph neural network has important theoretical and practical significance,and can be applied in the real system to obtain more efficient,accurate and robust results.