Research On Knowledge Graph Embedding Method For Link Predictio

Knowledge graphs structure the knowledge existing in the real world through entities and relationships,and store them in the form of graphs,with entities as nodes and relationships as edges.With the continuous expansion of knowledge the volume of the knowledge graph keeps increasing,the problem of incompleteness of the knowledge graph becomes increasingly prominent.Its incompleteness limits the downstream tasks of knowledge graphs,such as: recommendation systems,expert systems,etc.To solve the problem of incomplete graph information and better assist downstream applications,knowledge graph complementation tasks have emerged.The link prediction task,as a classical knowledge-completion task,not only completes the information of the knowledge graph,but also represents the entities and vectors in the graph as continuous vectors with low dimensionality,and the process is called knowledge graph embedding.Efficient knowledge graph embedding can accurately characterize the graph information and effectively improve the performance of downstream tasks.The development of deep learning has led to the development of knowledge graph embedding technology,and a large number of neural network-based models have been generated,which have significant advantages over traditional models.The knowledge graph not only has semantic features but also contains graph features.At present,the input processing of the graph data for the knowledge graph can be roughly divided into the processing of single triad information and the processing of the whole graph information.On the one hand,the feature information extraction of a single triad has room for improvement,and deep feature extraction can be carried out by building a deep network based on the current model.On the other hand,a single triad needs to rely on the global information of its atlas to maximize the feature information within the triad and the interaction information between the triads to further enhance the embedded information representation.Based on this analysis,the main research content of this paper is as follows.(1)A knowledge graph embedding model based on Multi-scale Covolution and Feature Assimilation Capsnet Embedding(MCACaps E)is proposed.The MCACaps E model is proposed for existing models that cannot fully capture the information inside the triad,and uses multi-scale convolution to expand the range of convolution layers to more comprehensively extract the features of entities and relationships inside the triad.The model extends the original capsule neural network by extending the convolution kernel of the original convolutional feature extraction layer to extract more information about the interactions between triads,and the corresponding features are performed in the feature map generation stage.The new feature matrix is obtained after feature fusion to generate the first layer of feature capsules,and the first layer of feature capsules is used to obtain the second layer of capsules through a dynamic routing mechanism.The second layer capsule weights are averaged to be the final vector,and the final ternary group confidence is obtained by the compression function,and the existence of the ternary group is judged based on the high or low score.Extensive experimental tests are conducted on the classical public datasets FB15k-237 and WN18 RR and the results are analyzed.The MR and H@10 metrics on the FB15k-237 dataset and the MRR and H@10 metrics on the WN18 RR dataset achieve the optimal values of the current comparison model.(2)The paper proposes a knowledge graph embedding model(Graph Neural Networks and Multi-scale Covolution and Feature Assimilation Capsnet Embedding,GNN-MCACaps E)by fusing graph neural networks and capsule neural networks.The GNN-MCACaps E model is proposed to address the problem that convolutional neural networks only process a single triad without considering the global information,and add the atlas information to the MCACaps E model.The model is improved in the following two aspects: first,the initial vector of Trans E of the capsule neural network is replaced by the vector obtained from the training of the graph neural network model.Second,in order to retain the graph information,the initial vector is input to the network,while the initial vector is added to the first layer of the capsule as a feature capsule,and the original weight averaging mechanism for generating the final capsule is improved to a weighting mechanism based on the attention mechanism.Firstly,the graph neural network is used to generate entity and relationship vectors containing global information,and the obtained vectors are input to the multi-scale convolutional capsule network to generate the corresponding global information capsules,at which time the first layer capsule includes the base triad capsule and the capsule of global information.The first layer capsule uses the dynamic routing mechanism algorithm to generate the second layer capsule,and the second layer capsule gets the final capsule by the weighting algorithm based on the attention mechanism,which dynamically fits the influence of each capsule feature on the final score.The global information is aggregated and each capsule information is dynamically matched to improve the accuracy of link prediction.The model experiments are tested in a large number of experiments on the WN18 RR and FB15k-237 public datasets and the results are analyzed.The MRR and H@10 metrics on the FB15k-237 dataset and the MR and H@10 metrics on the WN18 RR dataset are improved compared to the MCACaps E model.The MRR and H@10 metrics on the FB15k-237 dataset and the MR and H@10 metrics on the WN18 RR dataset were improved compared with the MCACapsE model.