Research On The Method And Application Of Heterogeneous Graph Embedding Based On Attribute Completion

Heterogeneous graphs can well describe objects and their interactions in the real world,and analyzing Heterogeneous graph is helpful to modeling the real world.However,the previous graph embedding methods are very sensitive to the node attributes of given heterogeneous graphs when analyzing them.At the same time,non-attribute nodes usually lead to noise propagation in the graphs,affecting subsequent downstream tasks.This makes the completion of node attributes of heterogeneous graphs an urgent problem to be solved.This paper studies this issue,mainly including the following three aspects:Firstly,because the existing methods of completing attributes through first-order neighbors can’t target the graph with randomly missing attributes,this paper proposes a heterogeneous graph attribute completion method and model based on attribute neighborhood.This method completes the non-attribute nodes in two stages.In the first stage,the attribute neighborhood is constructed by using the meta-paths.The attribute neighborhood aggregation can capture the semantic information of the attribute nodes and obtain the initial complement of the non-attribute nodes.In the second stage,the far and near relationship between nodes is obtained through structural information to further improve the attribute of completion.The node classification experiments on the open dataset have achieved more than 90% F1 value,which is obviously superior to the state-of-the-art attribute completion method of heterogeneous in all evaluation indicators.Secondly,in order to solve the problem that different types of nodes have no difference in the process of attribute neighborhood aggregation,this paper proposes a heterogeneous graph embedding method based on attribute completion.On the basis of the previous work,the hierarchical idea is used to build a multi-level homogeneous sub neighborhood of nodes,and the differences of node neighbor categories are captured by learning the sub neighborhood attributes of different categories.In addition,we further propose the completion attribute graph embedding module to solve the attribute completion problem of heterogeneous graphs with different attribute missing ratios,and optimize the graph embedding learning process.The node classification experiment on the public dataset shows that the proposed method performs better than the previous work in attribute completion,and the F1 value is increased by about 3%.Finally,this paper designs and implements a heterogeneous graph analysis system based on attribute completion.The system mainly realizes four functional modules: configuration module,analysis module,display module and management module.The configuration module provides task customization and parameter adjustment functions.The analysis module automatically completes heterogeneous graph attributes and embedding learning tasks according to tasks,and visualizes task results in the form of charts.The display module provides visual data display and historical task list functions.The management module provides heterogeneous graphs upload and complete attributes download functions.By implementing the above functional modules,it provides data analysts with the attribute completion of heterogeneous graphs and the visualization function of data mining.To sum up,this paper mainly aims at the problem that the existing heterogeneous graph embedding methods cannot accurately and efficiently learn the attribute missing heterogeneous graphs.The paper proposes the attribute completion method and model based on attribute neighborhood and the embedding method and model based on attribute completion heterogeneous graphs,which effectively complete the missing attributes in the heterogeneous graphs and optimize the end-to-end heterogeneous graph embedding method.Finally,a heterogeneous graph analysis system based on attribute completion is designed and implemented.The methods proposed in this paper have important theoretical significance and application value for the study of heterogeneous graph attribute completion and embedding.