Research On Graph Representation Learning Method Based On Contrastive Learning And Anchor Selection Strategy

Graph data is a kind of high-dimensional and abstract data composed of nodes and edges.It can model the data with complex structure in many different fields,such as drug target interaction data in the field of medicine,complex character relationship data in the field of social recommendation and so on.Traditional machine learning and deep learning methods can not be directly applied to high-dimensional and abstract graph data.Mapping high-dimensional and abstract graph data into low-dimensional feature vector representation through graph representation learning method has become a research hotspot.The low dimensional feature vector obtained by graph representation learning not only retains the topological information of the graph,but also can be widely used in many graph data analysis and processing tasks.However,the existing graph representation learning methods often pay attention to the overall information of the graph and ignore the differences between local information and similar information.Therefore,aiming at the correlation between captured information and the difference between similar information in graph neural network,this thesis proposes two graph representation learning models based on graph learning theory,comparative learning and anchor selection strategy.The main research contents are as follows:(1)A graph representation learning model of drug target network based on comparative learning is proposed,which enhances the correlation between node level embedded representation and graph level embedded representation in drug target network.Guided by the comparative learning paradigm,the model learns the enhanced node embedding representation by aggregating node information based on relationship perception and meta path.The encoder part is a graph convolution network,and the decoder restores the original heterogeneous network by inner product of the learned node representation.At the same time,the framework uses two levels of mutual information:(I)maximize the local mutual information,so that the learned node embedded representation captures the feature information of the whole graph;(II)Maximize the global mutual information,which is used to constrain the embedded representation of nodes to have the desired statistical characteristics.Finally,the drug target interaction link prediction task is carried out on the drug target correlation map network dataset.The experimental results show that the introduction of multi-level mutual information can effectively improve the prediction accuracy of the model.(2)A graph representation learning model based on anchor selection strategy is proposed,so that the nodes in the graph network can capture the global location information.The model uses the important nodes with a large amount of information interaction with other ordinary nodes in the graph as the anchor point,and filters the anchor points based on information entropy,and gives the calculation process of the distance from any given target node to each anchor point.Then,when initializing the node embedding representation,the relative distance information between each node and the other nodes in the graph is integrated.In addition,the model adopts the learning strategy of nonlinear distance weighted aggregation combined with the anchor set.In this way,the feature vector obtained by the given target node in the graph contains the position information of other nodes in the graph related to the selected anchor.Finally,the node classification task is carried out on the community discovery graph network dataset.The experimental results show that the anchor selection strategy can effectively improve the classification performance of the model.This thesis mainly studies the problem that the traditional graph neural network pays insufficient attention to the local information and the similar local information can not be well distinguished,and puts forward two schemes to improve the performance of the model,so as to better carry out the downstream graph network data analysis and processing tasks such as node classification and link prediction.