Learning Methods And Application Of Knowledge Graph Distributed Representation

As artificial intelligence gradually moves towards cognitive intelligence,knowl-edge graph has become an important form of knowledge representation and organiza-tion in the era of big data,and is widely used in fields such as semantic search,question answering,and personalized recommendations to empower their knowledge.Knowl-edge graph is composed of entities and relations.Entities represent real-world objects and relations express the relationships between entities.Traditional knowledge graph uses symbolic representation that each entity and relation are represented by unique strings.However,this kind of symbolism cannot meet the needs of large-scale knowl-edge graphs,cannot express the potential semantic connections between entities,and hinders the application of knowledge graphs.In recent years,with the development of knowledge graph technology,a distributed representation of knowledge graph has been proposed and used to solve the above problems.The distributed representation of knowledge graph maps entities and relations into a low-dimensional continuous vector space,and uses corresponding vectors to represent their semantic information.The dis-tributed representation of knowledge graph has the advantages of high efficiency and convenience,making the learning method and application of distributed representation a hot research topic in the field of knowledge graph at this stage.However,the widespread sparseness of entities in knowledge graphs makes it dif-ficult for the existing knowledge graph distributed representation learning technology to acquire high-quality distributed representations.At the same time,due to the hetero-geneity between the structured information of the knowledge graph and the user-item interaction data in the recommendation system,it becomes difficult to directly apply the distributed representation of the knowledge graph to the personalized recommendation task.After the in-depth investigation and analysis of existing distributed representation learning methods for knowledge graph,which attempts to solve the sparsity problem,and related research on recommendation algorithms combined with knowledge graphs,this dissertation carried out the following three research work in the learning methods and applications of knowledge graph distributed representation:First of all,this dissertation studies how to integrate the type information of en-tities into the existing knowledge graph distributed representation learning method to alleviate the problem of data sparsity,thereby improving the performance of distributed representation.Types not only restrict the entities belonging to the same type,but also contain two other important information,that is the hierarchical structure of types and the type constraint of relations.In order to model the antisymmetry and transitivity of the hierarchical structure,and use the type constraint of relations,we propose a new dis-tributed representation learning method combining hierarchical type information.First,we embed the types into a different vector space,then use the Order Embedding tech-nique to model the hierarchical structure of types on the type vector space,and then map entity embeddings to the type vector space through linear transformation to ensure that the entities and their types satisfy the partial order.For the type constraint of relations,we map the head(tail)entities corresponding to the relations to the type vector space in the same way,and ensure that the entities and the constrained types of the relations satisfy the partial order.In addition,even if the type constraint of relations information is unknown,our method can still exploit this information to guide the learning of dis-tributed representation.Based on four benchmark data sets,the experimental results on various tasks verify the superiority of our method.Secondly,this dissertation studies how to combine the logical information of soft rules with the existing knowledge graph distributed representation learning method to improve the performance of distributed representation.Soft rules are logical rules with confidence which are automatically extracted from knowledge graph.Considering that they are convenient to obtain and can support uncertainty,many researchers combine soft rules into distributed representation learning methods.However,the existing works either can not support the complex composition rules,or take soft rules as regularization terms to constrain derived facts,which is incapable of encoding the logical background knowledge about facts contained in soft rules.In addition,these works ignore forward chaining inference,which can further obtain more useful information.Therefore,we designed a learning algorithm for joint training over facts and soft rules.For modeling soft rules,fuzzy logic theory is used to model the groundings of rules generated by for-ward chaining inference,and soft rules are modeled by their corresponding groundings.In addition,in order to support massive rules and large-scale knowledge graphs for ef-ficient inference,we designed and implemented a distributed rule engine system based on the distributed memory computing platform Spark.Experiments on two large-scale knowledge graphs show the superiority of our joint training algorithm and the necessity of introducing forward chaining.F inally,this dissertation focuses on the study of applying distributed representation of knowledge graph in recommendation systems.Recommendation systems generally have the problems of sparse use-item interaction data,cold start and poor interpretabil-ity of recommendation results.The introduction of knowledge graphs can help allevi-ate these problems and obtain better personalized recommendations.The propagation-based methods in the existing related works have achieved the state-of-the-art perfor-mance,since they utilize both the information of entity and relation embeddings and the high-order connectivities within knowledge graph structure.However,these works either neglect the truth that different users may have different interests on entities and relations in knowledge graph or simply consider the different users' preferences over re-lations.Therefore,this dissertation proposes a personalized recommendation algorithm based on a fine-grained knowledge-graph-aware attention mechanism.In the process of propagating entity embeddings,which have high-order connectivity with items,we designed a fine-grained attention mechanism to generate weights for these entity embed-dings depending on the specific users.This attention mechanism in turn considers the user preference over the relation paths and associated entities.The experimental results in four real recommendation scenarios demonstrate the effectiveness and superiority of our method.