Research On Personalized Recommendation Algorithm And Application Based On Multi-source Heterogeneous Knowledge Graph

With the rapid development of the Internet,all kinds of data flooded various platforms,people enjoy the convenience of these data at the same time,the problem of "information overload" also comes along.The recommendation system is one of the effective ways to solve this problem and is also a current research hotspot in the field of machine learning.To cope with the problem of data sparseness and cold start in collaborative filtering recommendation systems,auxiliary information is usually used to improve the overall performance of recommendation systems.Currently,most social media websites and e-commerce systems allow users to post textual reviews to express their personal opinions towards the purchased items(e.g.,customers,movies,goods),along with a rating score indicating their preferences.In order to effectively integrate multiple data information,solve the problem of data sparseness,improve the accuracy of recommendation algorithms,a model with two components is constructed: one component is a model based on the fusion knowledge map of user-project history interactive data sources,and the other component is a deep learning algorithm based on user-project history commentary,which dynamically fuses the two components and solves the model using stochastic gradient descent method to provide users with more accurate personalized recommendation services.In this paper,we study the recommendation system based on multiple heterogeneous knowledge sources,modeling knowledge mapping and textual knowledge based on structured knowledge composition,respectively,with the aim of designing a recommendation system with superior performance.The main research contribution of this paper consists of the following three points.1)Traditional recommendation algorithms based on collaborative filtering have the problem of inefficient recommendation due to sparse data,which can be improved by adding supporting information;traditional recommendation algorithms based on knowledge mapping only focus on structured knowledge,however,user-item comment text knowledge is commonly found in various social networking sites and e-commerce systems,which contains rich semantic feature information,considering the features obtained from comment text knowledge can greatly improve recommendation efficiency.In this paper,knowledge mapping and comment content are treated as multivariate data elements,and different algorithms are used to process the data,and a dynamic fusion recommendation model,referred to as REME(Ripple Net and word2 vec fusion model),is proposed based on the historical user-item interaction information.2)Considering the complexity of commenting text knowledge and the need to extract valid feature vectors from it,this paper uses the deep learning model word2 vec to extract user and item related feature vectors from the text and calculate user-item similarity using the Factorization Machine(FM)algorithm to get a user click prediction value.3)In order to further verify the superiority of the REME algorithm,the experiments conducted on the real data set Yelp in this paper show that compared with the classical recommendation algorithm,using precision@k,recall@k and F1-measure@K as evaluation indicators,the REME model can achieve better recommendation effect and can effectively solve the problem of reduced recommendation accuracy due to data sparsity.