Entity Alignment Of Science And Technology Knowledge Graph Based On Node Importance Learning

In the construction of scientific and technological knowledge graph,entity alignment technology is needed to disambiguate the names of non-standard entities.However,because the same scientific and technological entity has different expression forms,or the same expression form will point to multiple scientific and technological entities,the accurate identification of scientific and technological entities is affected.In this thesis,the importance of nodes is used to align the graph of scientific and technological knowledge.The main research contents are as follows:(1)Node importance ranking of graph neural network based on compressed sensingFirstly,in view of the fact that the traditional node centrality measurement methods usually ignore the location of nodes,this thesis uses the framework based on compressed sensing to calculate the medium centrality,which is mainly to identify the high medium central nodes more accurately through indirect aggregation measurement without fully understanding the network topology;Secondly,for the general codec does not consider the node attribute,in order to solve this limitation,the codec method based on graph neural network is introduced.The encoder encodes the node as an embedded vector,and then maps the embedded vector to the decoder of BC scalar value calculated by compressed sensing.Experiments show that the proposed method,like DrBC method,uses deep learning method to capture the required features of nodes,but has more advantages in accuracy than DrBC method.(2)Entity alignment based on graph isomorphic network and important nodesEntity alignment is usually affected by structural heterogeneity and limited seed alignment problems.The heterogeneity of different knowledge graph(KG)structures is often very different,which may mislead the representation learning and alignment of seeds.The multi-channel graph neural network method is introduced,and different channels are used to complete KG self-attention,cross-KG attention and pruning.Generally weaker graph neural network variants are often unsuitable for training data.The graph isomorphic network used in this thesis has high representation ability,is almost suitable for training data,and is also superior to other graph neural networks in terms of test set accuracy;This thesis uses the importance of nodes for alignment.First find a class of nodes with greater importance,and then align these nodes in the alignment process.Experiments show that using node importance to align,its effect is better.(3)Scientific and technological knowledge atlas integration systemAiming at the heterogeneity and diversity of scientific and technological knowledge graph,an integrated system of scientific and technological knowledge graph is designed,which mainly includes data cleaning module and data integration module.Firstly,the scientific and technological data are cleaned and stored in the graph database,and then the nodes in the graph are sorted by importance,and then the top-N% nodes entered by the user are aligned.Finally,the alignment results of knowledge graph are returned to the user...