Research On Robust Graph Neural Networks For Attributed Networks Analysis

Attributed networks are prevalent in the real world,such as citation networks,biological networks,and sensor networks.Graph neural network(GNN),as important tools for attributed network analysis,possesses various advantages including expressive power,generalization ability,and transferability.It has been widely used in link prediction,node classification,attribute completion and other tasks.Most GNN models assume that attributed networks are clean and reliable.However,many attributed networks do not meet this ideal assumption,leading to inadequate robustness of existing GNN models on class-imbalanced,noisy,and incomplete attributed networks.Given the ubiquitous presence of these three types of attributed networks,purposefully modeling GNN models to enhance their robustness to these attributed networks holds significant theoretical and practical significance.Due to the complexity of attributed network scenarios,there are many unresolved issues in developing robust GNN models for these three types of attributed networks.Therefore,developing robust GNN models remains a challenging problem.Based on the above,this thesis concentrates on robust GNN research for the prevalent and representative class-imbalanced,noisy,and incomplete attributed networks.And focus on the following three problems:(1)For class-imbalanced attributed networks,most GNN models employ oversampling techniques to alleviate the problem of class-imbalanced in graphs.However,existing oversampling-based GNN models overlook the problem where the representations of minority nodes are dominated by majority nodes due to the aggregation of neighbor information by GNN.(2)For noisy attributed networks,the variational graph autoencoder(VGAE)represents a promising class of GNN models for handling noisy attributed networks.However,the VGAE,which relies on the mean-field assumption,may not effectively represent noisy attributed networks,and employing only one variational network may limit the model’s ability to learn richer network information.(3)For incomplete attributed networks,GNN models based on data imputation techniques offer an effective approach for dealing with incomplete attributed networks,where the structure is complete but some node’s attributes are missing(referred to as attribute-missing attributed networks).However,existing GNN models based on data imputation face challenges related to inaccurate input and insufficient expressiveness of decoders.To address the aforementioned problems,this thesis leverages GNN-related theories to propose GNN models specifically designed for class-imbalanced,noisy,and incomplete attributed networks,with the goal of enhancing the robustness of GNN models.The primary contributions and innovations are outlined below:(1)To address the problem of minority nodes being dominated by majority nodes in existing GNN models based on oversampling techniques,this thesis introduces a novel graph oversampling framework based on distribution alignment.The aim is to enhance the robustness of GNN models to class-imbalanced attributed networks.The proposed GNN model comprises two key components: a graph autoencoder module based on a multilayer perceptron and a distribution alignment module based on a sumproduct network(SPN).The graph autoencoder module with a multilayer perceptron is employed to generate synthetic minority nodes before aggregating neighbor information by GNN,mitigating the dominance issue of majority nodes resulting from GNN’s message propagation.The distribution alignment module based on SPN utilizes distribution alignment method to gather more information about minority nodes,alleviating distribution mismatch problem caused by aggregating neighbor information through GNN.This thesis introduces SPN for the first time to address class-imbalanced attributed networks.Experimental results on class-imbalanced datasets demonstrate the superiority of the proposed GNN model for class-imbalanced attributed networks over contemporaneous comparative algorithms,showcasing increased robustness in handling class-imbalanced attributed networks.When the imbalance ratio is 0.1,the effect of the proposed GNN model is more significant.(2)In response to the problem of the VGAE based on the mean-field assumption not adapting well to noisy attributed networks,a multi-head VGAE model constrained by SPN is proposed to enhance the robustness of GNN models to noisy attributed networks.The proposed GNN model comprises two key components: a multi-head VGAE module and an SPN-constrained module.The multi-head VGAE module learns richer attribute and structural information through multiple variational networks.The SPN-constrained module relaxes the mean-field assumption using SPN to model more complex and robust probability distributions,thereby improving the model’s robustness.This thesis integrates SPN with GNN for the first time,enhancing GNN’s robustness to noisy attributed networks and extending the application scope of SPN.Experimental results on five benchmark graph datasets demonstrate that the proposed GNN model for noisy attributed networks outperforms contemporaneous comparative algorithms,exhibiting increased robustness to noisy attributed networks.(3)To address the problem of inaccurate input and insufficient expressiveness of decoders in existing GNN models based on data imputation,a novel attributeimputation graph autoencoder tailored for attribute-missing attributed networks is proposed.It aims to enhance the robustness of GNN models when dealing with attribute-missing attributed networks.The proposed GNN model comprises three key components: a dual encoder,a data imputation module,and a dual decoder.Specifically,during the encoding phase,a dual encoder based on knowledge distillation is devised to simultaneously encode attribute and structural information into the representation of attribute-missing nodes,facilitating the acquisition of representations for attributemissing nodes.The data imputation module,by recombining representations of attribute-missing and attribute-observed nodes,avoids the introduction of noisy information.In the decoding phase,a multi-scale masked decoder is designed to enhance the expressiveness of the decoder through the incorporation of multi-scale and masking modules.Experimental results on commonly used datasets for incomplete attributed networks demonstrate that the proposed GNN model tailored for attributemissing attributed networks outperforms comparative algorithms,exhibiting increased robustness to attribute-missing attributed networks.When the attribute missing ratio is0.8,the effect of the proposed GNN model is more significant.In summary,this thesis proposes three robust GNN models for attributed network analysis,aiming to address the problems of class imbalance,noise,and incompleteness commonly encountered in attributed networks.To address class imbalance and noise problems in attributed networks,the SPN method is introduced to improved the GNN model.To tackle the problem of incomplete attributed networks,masking and multiscale techniques are introduced to improve the GNN model.These three proposed models are all built upon the foundation of graph autoencoders,effectively modeling various types of attributed networks.The effectiveness of the proposed robust GNN models is validated through theoretical derivations and experimental analyses.It is hoped that the models proposed in this thesis can enrich research in the field of robust GNNs and provide inspiration and advancement for the development of robust GNNs for attributed network analysis.Furthermore,this thesis extensively discusses the current challenges faced by robust GNN research targeting different attributed networks and offers insights into future research trends in this domain.