Enhancement And Optimization Of Graph Neural Networks

Graphs,serving as fundamental data structures,represent relationships between entities and have widespread applications in diverse fields,including biology,chemistry,finance,and education.Their adaptability is crucial in a variety of artificial intelligence applications.For example,leveraging graph structures enables the prediction of molecular chemical properties,the detection of financial fraud,and the generation of friend recommendations on social media platforms.To capitalize on this graph data,Graph Neural Networks(GNNs)have been developed.These networks expand the capabilities of traditional deep neural networks(DNNs),moving from grid-based data like images or text to structured data.This advancement allows for the efficient processing of non-uniform graph-based data.GNNs have shown consistently superior performance in various graph tasks,such as node classification,graph classification,and link prediction.However,increasing evidence suggests that GNNs have limitations.Over-reliance on the message-passing framework can induce excessive similarity among node features.Moreover,over-correlations between distinct feature dimensions might lead to the omission of crucial information,subsequently undermining GNNs’ efficacy.Moreover,GNNs also inherit the vulnerabilities of DNNs,especially when the data is subjected to adversarial perturbations.Even small perturbations to node features or the graph structure can substantially hamper GNNs’ performance.Such adversarial perturbations can have profound implications in contexts like financial fraud detection or medical recommendations.Hence,fortifying the robustness of GNNs has become a priority in both academic and industrial spheres,underscoring the quest for safety and dependability across these sectors.Although GNNs have notably succeeded with static unsigned networks,many realworld graphs are dynamic and possess signed attributes.Temporal Signed Networks,encapsulating these dual traits,are prevalent across diverse applications.These networks enrich our dataset with augmented informational depth and time-based perspectives,facilitating a more nuanced understanding of intricate systems.Investigating these networks not only furnishes a foundation for informed decision-making across various disciplines but also introduces fresh avenues and intricacies in GNN research.Considering the challenges faced by GNNs,this thesis proposes three principal lines of research:First,this thesis investigates the pervasive issue of feature over-correlation in GNNs and finds that the parameter update mechanisms in existing models often fail to maintain independence among features.Through rigorous empirical research and in-depth theoretical analysis,we ascertain that the feature propagation process can lead to high correlations between features.To address this issue,we introduce a novel propagation strategy named "DeProp",and validate its high compatibility with existing GNN frameworks.Experimental results confirm that DeProp not only significantly enhances the performance of deep GNNs but also resolves issues of over-correlation and oversmoothing.Second,this thesis introduces a new method aimed at countering adversarial attacks on graph data through a structured learning approach based on elastic GNNs.This method initially employs elastic GNNs to detect and identify anomalous edges,followed by denoising and structured learning of the graph data.The presented elastic message estimator,incorporating different types of norms,effectively performs denoising and addresses the issue of non-smoothness in anomalous edges.Numerical experiments show that this strategy significantly enhances the learning performance and robustness in handling graph data under adversarial attacks and noise interference.Last,this thesis conducts a detailed analysis of the structural balance evolution in temporal signed networks and selects a Twitter dataset,rich in unfollow information,to examine user unfollow behavior.To this end,we design a temporal signed GNN model,aimed at accurately predicting user unfollow actions.Experimental results confirm that our method performs well in predicting unfollow behavior,validating the effectiveness of the proposed model in handling applications within temporal signed networksOverall,this thesis aims to address the limitations of GNNs from both technical and application perspectives.Specifically,this thesis introduces a decorrelation propagation scheme,a structured learning approach based on elastic GNNs,and a temporal signed GNN model,all of which are designed to enhance and optimize the performance of GNNs.