Research On Deep Graph Encoding Feature Learning Algorithm In Brain Disease Data Analysis

Resting-state functional Magnetic Resonance Imaging(rs-f MRI)data is complex and the feature dimensions of extracted brain functional connectivity data is often high-dimensional,which is not conducive to disease identification.Since there is similarity between subjects(such as gender,age,and region),the similarity of all subjects is captured by constructing graph structure.Currently,most researchers commonly use deep feature learning methods to reduce the dimensionality of brain disease data and extract more representative low-dimensional feature information.However,in these studies,label information is often used for disease identification,which is practically costly.To address the above problems,this thesis mainly proposes two unsupervised feature learning methods based on deep graph encoding to extract brain functional connectivity data features.The details of the study are as follows:First,Graph Attention Auto-encoder with Self-representation feature selection(GAAS)is proposed,which is an unsupervised method to mine low-dimensional feature representations.Graph attention encoder captures the similarity between nodes,focuses on the local neighborhood information of nodes.Therefore,from the feature subspace perspective,this method designs a feature self-representation module to enhance the local-to-global linear information.Then,a graph constraint term is exploited to enhance the similarity of adjacent nodes in the original graph,and the importance representation of neighboring nodes in the self-representation module is constrained in the form of a graph structure.In addition,the self-representation coefficient matrix that representing the global feature relationships is used as the criteria to judge the importance of features for feature selection.Finally,the most discriminative feature subset of brain disease subtypes is selected in descending order of importance for brain disease recognition.Extensive experimental results on eight brain functional connectivity datasets demonstrate that the selected feature subsets can accurately identify brain disease subtypes,and the proposed method has highly competitive among existing feature learning methods.Second,on the basis of the first work,graph contrastive representation learning is incorporated,and the original data is expanded through data enhancement.Human brain disease data is complex,and multiple disease subtypes may exist for the same disease.Existing graph contrastive learning methods mainly consider only the representation information of individual diseases for contrast,without identifying the similarity information between individuals.This may affect the quality of information node representation in the graph.To this end,this thesis proposes a method named Joint Graph Contrastive and Generative Representation Learning(GCGRL).Firstly,a novel graph contrastive method is designed to learn high-quality consistent representations by constructing multiple positive and negative sample pairs.In addition,a constraint mechanism is designed,which considers the graph topological reconstruction in the corrupted data to enhance the similarity between node and neighborhood embeddings.Extensive experimental results on eight brain functional connectivity datasets prove that GCGRL improves similarity between individual representations and the accuracy of brain diseases recognition.Compared with the existing state-of-the-art methods,the results show the effectiveness of proposed method.