Research On Knowledge Hypergraph Link Prediction Based On Attention Convolution Network And Residual Perceptron

Knowledge hypergraph Link Prediction(KHGLP)is one of the effective means to complement the knowledge hypergraph by predicting the unknown n-ary relations from the known entities and n-ary relations in the knowledge hypergraph.It can be used in many fields such as question answering systems,semantic analysis,and so on.However,the existing optimal embedding model-based knowledge hypergraph link prediction method,Hyp E,although the entity embedding considers the location,directly uses the initialized relation embedding to participate in the final scoring,without considering the different degrees of contribution of different entities to the relation,and the entity embedding does not contain sufficient information,thus limiting the algorithm performance.Therefore,we have carried out research,and the main research work includes:(1)To address the above problems,we propose a knowledge hypergraph link prediction method LPACN(Link Prediction based on Attention Convolution Network)based on attention convolution network.First,the method uses the attention mechanism to solve the problem of entity contribution to the relation,and the attention mechanism is used to process the tuple to extract the information in the entity that is beneficial to the relation according to the attention weights and add this information to the relation vector.Then,for the problem that the entity embedding does not contain enough information,we add information about the number of neighboring entities within the same tuple to the convolutional kernel that extracts entity features in the convolutional network,so that the entity embedding contains more information.(2)Although the performance of LPACN is improved,there is still the problem of gradient vanishing,so we continue to propose an improved method LPACN~+based on residual perceptron.To alleviate the gradient vanishing problem,we use an improved Residual Net to optimize LPACN by a hopping connection to recover the vanished gradients for the model and uses a soft threshold module to improve the robustness.In addition,to further enhance the nonlinear learning capability of the model,a Multilayer Perceptron(MLP)is added after the Residual Net to obtain the final improved model LPACN~+.The results of numerous experiments conducted on real datasets show that the LPACN method we proposed improves the link prediction performance by 7.4%compared to the baseline methods.The results of comparison experiments also verify the effectiveness of the improved model LPACN~+in this thesis.