Graph Embedding Method Based On The Biased Walk On Neighborhood Attributes

Graph embedding,also known as graph representation learning,is a method of mapping high-dimensional graph data into low-dimensional vectors,which aims to solve the problem that graph-structured data is difficult to input into the machine learning algorithms efficiently.The essence of random walk is a finite Markov chain of time exchange.Applying random walk to graph embedding can apply many classical high-dimensional data algorithms to link prediction,node classification,and other tasks.The random walk can perform graph representation learning based on topological similarity,improve the expressiveness of data,and reduce the complexity of graph data.The random walk method is improved by introducing the idea of combining depth-first and breadth-first,these improvements help random walks to capture the attribute information of graphs and improve graph embedding.Existing studies have unsatisfactory effects on the selection of neighbor nodes in the process of graph embedding,and it is difficult to effectively improve the accuracy of graph embedding.The selected neighbor nodes often have poor effects especially when the data set contains many scattered subgraphs.This thesis proposes a graph embedding method to obtain a better embedding effect by improving the quality of random walks.The random walks used in graph embeddings are improved from the graphs' attribute information and topology structure.The main work of the thesis includes:(1)Graph embedding method based on multi-class random walks of neighborhood attributesNode classification is performed on the data,and nodes are classified according to their attributes or the similarity between nodes.During random walks,different types of nodes are selected.The quality of random walks selection of nodes is improved by selecting different types of nodes.By classifying the attributes of neighboring nodes,the attribute information of the graph can be captured by random walks.A better embedding effect can be obtained by using the breadth-first idea in the neighborhood aggregation process.The idea of node embedding is introduced into the edge embedding,and the embedding result of the edge is obtained.(2)Graph embedding method with biased walk based on neighborhood attributesBased on the method in the previous section,data enhancement is added,and the node walking strategy is optimized through the data enhancement method.The idea of combining breadth-first and depth-first is applied to the process of node selection to improve the stability of the embedding effect.Migrate the parameter settings of supervised datasets to semisupervised datasets to improve the embedding effect of semi-supervised datasets.(3)Experimental verification and analysisThe method proposed in this thesis is experimentally validated and analyzed using four datasets.The results show that in different downstream tasks including node classification and link prediction,the algorithm in this thesis can effectively improve the effect of graph embedding.