Research And Application Of Heterogeneous Graph Network Algorithms

Heterogeneous graphs are suitable for describing the real-world data with different types of entities and relationships,effectively extending the concept of networks.As a novel technique to capture heterogeneous information in heterogeneous graphs,heterogeneous graph neural networks can effectively combine the message passing mechanism with the complex semantics in heterogeneous graphs and achieve great success in the field of heterogeneous graph mining.However,there are some generic problems in heterogeneous graphs,such as the problem of no initial features and the long-tail problem of node degree.Knowledge graphs are a special case of heterogeneous graphs,and graph modeling and graph representation methods are proposed based on both but rarely intersect,which is a direction of interest.A patent knowledge graph can model the interaction between a large number of patents,and the analysis of a patent knowledge graph using heterogeneous graph technology also faces generic problems such as missing patent features and long tails.To address the above challenges,this paper conducts an in-depth study on the key techniques of heterogeneous graph neural networks.First,two key generic problems of heterogeneous graphs,i.e.,feature missing and long-tail problem,are studied,and the two corresponding heterogeneous graph neural networks are designed to solve them respectively.Secondly,a patent knowledge graph construction method incorporating the semantics of heterogeneous graphs is investigated.Finally,the application of heterogeneous graph neural networks in patent knowledge mapping is further explored.In summary,the main research contents and innovation points of this paper are shown as follows:(1)A heterogeneous graph neural network model,named Position Encoding for Heterogeneous Graph Neural Network(PE),is proposed for the problem of how to generate high-quality features for heterogeneous graphs without initial features or difficult to generate features.Specifically,the topological embedding of nodes is obtained using graph embedding,the topological relationship between nodes is used as a guide to calculate the positions between nodes in a subgraph,and the position information is encoded as features as initial features or additional features for subsequent node aggregation.This study is a generic framework for heterogeneous graphs that can be easily combined with existing excellent heterogeneous aggregation models.Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed heterogeneous graph neural network framework.(2)A heterogeneous graph neural network model,named Heterogeneous Graph Neural Network with Tail Node Completion(HGNN-TC),is designed to address the long tail problem of node degrees in heterogeneous graphs.Specifically,node content transformation is performed to project heterogeneous nodes into the same feature space,and then the rich heterogeneous neighborhoods of the target head node are learned to generate global relationships,and the head and tail nodes are compared to generate additional information for tail node aggregation.This study also serves as a general framework for heterogeneous graphs that is easy to combine with arbitrary heterogeneous aggregation models.Extensive experiments on two benchmark datasets demonstrate the effectiveness of the proposed heterogeneous graph neural network model for the tail node classification task.(3)A patent knowledge graph construction scheme incorporating heterogeneous semantics is proposed to address the massive attributes and silo characteristics of patent data.Firstly,the patent data collected from the patent database is cleaned in multiple steps.Secondly,heterogeneous graph semantic rules are introduced,and the ontology of patent knowledge graph is designed with patents as the center.Based on the defined ontology,the cleaned data are mapped into entities and relationships and stored in the graph database to realize the construction of the patent knowledge graph.Finally,the patent heterogeneous graph is extracted based on the constructed patent knowledge graph,and the proposed heterogeneous graph neural network is used to perform efficient and highly accurate classification applications for the patents in the patent knowledge graph.