| Most knowledge graph suffers from incompleteness,to tackle with this problem,the knowledge graph completion technology is proposed,which can learn the pattern of graph structure,and realize the extrapolation of latent knowledges as well as the maintenance of the completeness of the knowledge graph automatically.The knowledges of the knowledge graph are from real world,and new knowledges can be generated with the change of the world,which means that the number of nodes in the graph can grow dynamically.Traditional knowledge graph completion methods are mostly trained on whole graph,the cost of such training strategy is hard to accept for graphs of large scale,thus how to perform link prediction on newly added nodes brought by the constant expansion of graph becomes an important topic.To deal with this problem,this thesis obtains the feature representation of the new nodes by aggregating its local neighbor structure,so as to realize the link prediction of the new nodes,the main contribution of this work is as follows:1.A knowledge graph completion model based on aggregating local structure information is proposed.Aiming at the problem that the existing methods are not comprehensive enough in the extraction and aggregation of local structure information,the comprehensiveness of structural information extraction is improved by improving the subgraph extraction strategy,and the sufficiency of information extraction is improved by aggregating local subgraph structural information based on the idea of tree-LSTM.Comparative experiments with existing models on standard datasets show that the proposed model has great improvement on both AUC-PR and HIT@10.Compare with the suboptimal results,the AUC-PR of the proposed method on FB15k-237-V3,NELL-995-V1 and NELL-995-V4 is improved by 2.50%,2.37% and 1.89%,respectively.2.A knowledge graph completion model based on aggregating global structure information of anchor node is proposed.Aiming at the problems of existing anchor node sampling method,a new anchor node sampling method based on graph segmentation algorithm is proposed,which can improve the uniformity of the distribution of the sampled anchor node set.Aiming at the problem of insufficient utilization of anchor node information in the previous anchor node-based method,a new method that can realize path information encoding based on subgraph aggregation is proposed.Compared with the existing methods,the method proposed in this thesis effectively improved the expression ability of new nodes,the proposed method got best HIT@10score on both o FB15k-237 and o WN18 RR dataset.3.Designed and implemented a knowledge graph system which is about the knowledge of academic,this system can update its knowledge incrementally by collecting data in the real world,and the two methods aforementioned are integrated into one module to perform completion of new nodes,there is also a front-end module for handling requests from users and presenting feedbacks. |
