Research On Large-scale Knowledge Graph Representation Learning Algorithms And Applications

With the rapid growth of information,how to make computers mine useful information and derive more valuable human knowledge from it is the main problem of research in artificial intelligence.The knowledge graph is a fundamental analysis tool for studying cognitive intelligence in the field of artificial intelligence,and its powerful reasoning power can be used to empower machines to learn human knowledge in the real world.Traditional knowledge graphs are usually stored using a triplet approach,where entities and relations in the triplet are represented symbolically.However,it is difficult to apply the symbolic representation to large-scale knowledge graphs.In order to address the problem,knowledge graph representation learning techniques have become a hot research topic in recent years.In addition,the knowledge graph provides an effective solution to the problems of unexplained,cold start,and sparsity of existing recommendation algorithms.Hence,we focus on the knowledge graph representation learning technique and its application on recommendation algorithms in this paper,and the main contents of the research are as follows:(1)To tackle the problem that the existing knowledge graph representation learning model does not consider the triplet scoring range,we design a method to calculate the joint cosine similarity of complex vectors and use it to improve the scoring function of the rotation model so that the triplet scoring range is bounded,thus reducing the model variance and improving the model expressiveness.At the same time,we introduce a triplet scoring range scaling factor to improve the loss function,which is used to adjust the triplet scoring range and facilitate the model training.In addition,since our method incorporates the rotational nature of the rotation model,it can model and reason about three important relational patterns of the knowledge graph.Experimental results show that our method improves the rotational model Hits@1 metric by 4.0% compared to the WN18 RR dataset and achieves some of the current state-of-the-art results.(2)To tackle the problem of insufficient feature representation capability of the deep knowledge-aware network model,we optimize the following two aspects: we use the previously proposed knowledge graph representation learning model for learning the vector representation of entities to mine deeper semantic information,to optimize the entity representation features;we propose a deep average network based on multisource information aggregation for learning the vector representation of sentences,and fuse information from words,entities,and contexts by designing different aggregation functions,to optimize the sentence representation features.Experimental results show that our approach improves the AUC metric by 5.4% in the Bing News news dataset compared to the deep knowledge-aware network model.