Deep Graph Learning And Applications

Graph data contains relationship and structural information among entities,and is widely used in real-life applications,such as social networks,knowledge graphs,and brain functional connectivity graphs.Despite the widespread use of graph data,its complexity and diversity pose significant challenges for intelligent perception and reasoning.With the advancements in deep learning,machine learning methods specifically designed for graph data have witnessed remarkable progress.Recently,graph convolutional neural networks(GCNs)have emerged as one of the prominent methods in the field of deep graph learning.Deep graph learning leverages the powerful learning capabilities of deep neural networks to extract structural information from graph data,to output discriminative representation and enables efficient and accurate analysis of graph data.Consequently,deep graph learning has been widely adopted in various domains,including social networks,recommendation systems,biomedicine,and network security.Despite the widespread applications of deep graph learning in various domains,there remain significant challenges that need to be addressed,particularly in practical scenarios involving complex graph data.For instance,real-world applications often involve lowquality graph data,unlabeled graph data,multiplex graph data,and multi-center medical graph data.Those data pose significant challenges in terms of the model’s robustness,generalization,efficiency,and privacy protection.To address these issues,this dissertation focuses on studying the complex graph data scenarios faced by deep graph learning in practical applications and proposes a series of high-robustness,high-efficiency,and highgeneralization deep graph learning methods.These methods aim to effectively bridge the gap between deep graph learning and practical applications.The primary contributions of this dissertation include:(1)Reverse graph learning for dynamic graph structural learning.In practical applications,the quality of graph structures is often challenging to guarantee,but it plays a crucial role in the performance of deep graph learning models.Constructing graph structures typically relies on the similarity of features in the raw data or manually defined rules.However,the presence of noise and outliers in the raw data and manually defined rules can lead to erroneous connections or lack of complete links,thus affecting the accuracy and robustness of the models.Additionally,most deep graph learning methods assume that the initial graph structure is of high quality and keeps it fixed during the training process.Consequently,a low-quality initial graph can adversely affect the representation learning process and limit the effectiveness of deep graph learning models in practical applications.To address these issues,this dissertation proposes a dynamic graph structure learning model based on reverse graph learning.By leveraging reverse graph learning,the model dynamically adjusts the initial graph structure to improve its quality and accuracy.Experimental results demonstrate that the proposed model enhances the robustness of deep graph learning and efficiently infers out-of-sample nodes,effectively addressing the challenges posed by low-quality initial graphs.(2)Contrastive graph learning with constraints.In practical applications,data acquisition is often relatively straightforward,while data annotation requires significant manual efforts.Therefore,deep graph learning needs to achieve effective representation learning for unlabeled graph data.Unsupervised graph representation learning methods based on contrastive learning have emerged as a promising approach to address this challenge.However,existing methods often overlook the gap between representation learning and downstream tasks when designing contrastive graph learning methods,limiting the generalization performance of the learned representations.Moreover,current studies often rely on empirical knowledge to design contrastive graph learning models,lacking systematic theoretical guidance for mathematically designing models.To address these issues,this dissertation proposes a contrastive graph learning model with constraints.The proposed model efficiently and robustly learns representations from unlabeled graph data by maximizing the mutual information between semantic information and structural information in the graph.Furthermore,a theoretical framework is established to explore the relationship between contrastive learning and downstream classification tasks.The proposed model introduces constraints specific to contrastive graph learning models to reduce the gap between representation learning and downstream tasks.Experimental results demonstrate that the proposed model exhibits efficient representation performance and remarkable generalization capabilities.(3)Multiplex graph learning with complementary and consistent information.In practical applications,graph data often exhibit multiple types of relationships among samples.For example,in social networks,there can be different types of relationships such as friendship and interest connections,which can be represented by multiple graph structures(multiplex graph).Deep graph learning models tailored to multiplex graph have significant research and application value as they can extract more comprehensive information from various relationships between nodes.However,existing methods face challenges such as noise and efficiency issues due to the diversity and complexity of multiplex graph.To address these challenges,this dissertation proposes a multiplex graph representation learning model based on complementary and consistent information.The model efficiently explores the complementary and consistent information in multiplex graph through local structure preservation and correlation constraints.Experimental results demonstrate that the proposed method achieves higher efficiency and stronger robustness against noise compared to existing methods.(4)Federated graph learning for early diagnosis of neurological disorders.According to statistics from the World Health Organization,the global incidence of neurological disorders is approximately 17%,which has become an urgent issue to be addressed in China and worldwide.Early diagnosis of neurological disorders is crucial for delaying disease progression and improving the quality of patients’ life due to the irreversibility of neuronal damage.Although deep graph learning models have been widely applied in various fields,few researchers have explored their application in the early diagnosis of neurological disorders.The challenge lies in the incomplete graph structure information derived from multi-center privacy-protected data and multi-center graph data.To address these challenges,this dissertation proposes a federated learning based graph learning model.The model leverages both clinical information and medical image information to construct the graph structure and incorporates a federated subgraph completion module to effectively handle local deviations from the multi-center models.Experimental results demonstrate that the proposed model exhibits excellent diagnostic performance in the early diagnosis of neurological disorders while preserving data privacy.Through the aforementioned methods,this dissertation provides an in-depth exploration of the technical challenges encountered by deep graph learning in practical applications and proposes corresponding solutions.As a result,this research has expanded the scope of the application of deep graph learning models and advanced the development of the deep graph learning field.