Research Of Knowledge Graph Embedding Adversarial Learning Method Based On Attention Mechanism

In recent years,many applications represent data in the form of graphs,especially Knowledge Graphs(KGs),which have received extensive attention due to their structured characteristics.KGs are structured representations of real-world information,which can model complex data in a machine-readable way,and contains rich knowledge,so it is widely applied to various domains,from question answering,to finance and education.However,KGs are usually huge and sparse,and suffer from incompleteness.Besides,the potential symbolic nature of such triples usually makes KGs hard to operate.To solve these problems,Knowledge Graph Embedding(KGE)has been proposed and developed rapidly.Knowledge Graph Embedding aims to learn the distributed representation of KGs,using lowdimensional dense vectors to represent entities and relations,which can simplify operations while maintaining the inherent structure.At present,a large number of KGE models have been proposed,among which the KGE models based on deep learning greatly improve the expression ability of embedded representation by using deep network structure.But they usually emphasize entity embedding,while the learning of relation embedding is relatively simple.And they only use triples,ignoring the rich inference information contained in the multi-hop paths.In addition,complex network structures are prone to over-fitting,which leads to their poor generalization ability on real datasets.In order to obtain high-quality and more robust embedding,this thesis proposes a knowledge graph embedding adversarial learning algorithm based on attention mechanism.The main work is as follows:(1)In view of the problems that the previous models cannot make full use of the potential information in KGs,and the relation embedding is relatively simple,a novel KGE approach called HAPKE(Hierarchical Attention with Relation Paths for Knowledge Graph Embedding)is proposed.HAPKE constructs a two-level attention encoder,which combined with relation paths to assist model learning embedding on the basis of deep learning.Firstly,at the triples-level,HAPKE utilizes the attention mechanism to learn the information on the triples and its neighborhood.Then at the paths-level,paths are filtered according to the semantic similarity between paths and relations,and paths are modeled based on the triples-level relation embedding for further updating the relation representation.Next,the embedded representation learned by the encoder is fed into the decoder,which extracts the latent features inside the triples and paths,and better maintains the translation property of the triples.Finally,the link prediction experiment is performed on four general KGs of FB15 K,FB15K-237,WN18 RR and Kinship,and case analysis is conducted on the UMLS dataset to verify the effectiveness of the HAPKE model.(2)Aiming at the problem that complex model structure is prone to over-fitting and leads to poor robustness of the model,a knowledge graph embedding model based on generative adversarial networks GAKGE is further proposed.Adversarial training can effectively enhance the robustness,so GAKGE introduces the generative adversarial network as an adversarial training component,which can essentially be regarded as a regularization term of the model.In fact,adversarial training can constraint the process of embedding learning and model the uncertainty of the data,which effectively alleviates the over-fitting problem and improves the generalization ability of the model,making the learned embedding representation more robust.Finally,link prediction experiment is carried out on four benchmark datasets,as well as a series of parameter sensitivity analysis and ablation experiments to verify the effectiveness of the GAKGE model.