Research On The Recommendation Algorithm Based On Knowledge Graph Representation Learning

In recent years,with the rapid development of the mobile Internet,the information on the Internet has become more and more abundant.In the face of such rich network information,recommendation is particularly important.The main function of the recommendation system is to mine the user's behavior preference according to the user's historical behavior information.Then according to the user's preference model to provide the user to meet the requirements of the information.The sparsity problem and cold boot problem of traditional recommendation system limit the effect of recommendation to some extent.However,representation learning realizes the projection of the rich correlation information between different research objects into the vector space of low dimension.This learning method can well solve the problems in the recommendation system.The combination of Knowledge graph representaton learning and recommendation can effectively improve the performance of recommendation and realize personalized recommendation in line with users' preferences.In the existing recommendation of knowledge graph based representation learning,translatation-based representation learning algorithm has achieved great success.However,in these methods,the structural information of the user and the item itself,as well as the interaction information between the user and the item,is not fully considered,leading to the poor effect of recommendation.Therefore,in view of the above problems,this paper has made relevant researches and improvements on translation-based presentation learning,mainly including the following aspects:1.Based on the analysis of entity structure information in the knowledge graph,a representation learning algorithm(Strans)which integrates entity structure information is proposed to achieve more accurate representation of entities and relationships in the knowledge graph.Specifically,we analyze the structure information of the entity,including direct structure information and indirect structure information.By learning the two parts of the entity structure information,the entity vector representation containing rich structure information is obtained.Based on the typical translation models,a new Strans E model and a Strans H model are obtained.Through a large number of experiments,the results show that Strans E and Strans H models have made significant progress in the task of knowledge graph linking prediction.2.From the perspective of analyzing the user's behavior information on the project.Existing presentation learning algorithms ignore the interaction information between users and dislike items,and their preferences do not achieve comprehensive learning.To solve this problem,this paper proposes a recommendation method based on user adaptive preference.Specifically,we first increase the number of relationships between users and projects in the existing classical translation model to train the "preference" relationship vectors of users.Secondly,we put forward the shortcomings of the algorithm in the training process and the corresponding solutions.Finally,we make full use of the negative feedback information to update the recommendation list.