Research On Personalized Recommendation Method For Multi-Source Heterogeneous Information Network

With the continuous development and integration of the Internet industry and real-world businesses,various cross-platform and cross-industry innovative service models have been born to meet the growing personalized needs of users.As a result,the network data scale is increasing and the data forms are gradually diversified.As a traditional means to solve information overload,recommendation system also needs to face the huge challenges brought by the reform of information industry.The current recommendation system gradually shows many new features.With the development of mobile network,the recommendation system should be universal and personalized;Due to the influence of multiple factors on the recommendation system,the recommendation structure is becoming more and more complex.How to implement recommendation to meet user needs in multi-source heterogeneous environment has also become a hot issue in the field of personalized recommendation.Heterogeneous information network,as a network carrying a large number of multi-type data and relationships,can help the recommendation system analyze and mine the potential connections hidden among various influencing factors,and improve the effectiveness and accuracy of recommendation.Therefore,studying the structural characteristics of heterogeneous networks and mining multiple potential relationships are of great significance to the current real and complex personalized recommendations.The main work of this paper is as follows:(1)Aiming at the problems of data sparsity and integrating multiple influencing factors in recommendation system,a hybrid recommendation model combining deep matrix factorization and tensor factorization is proposed.Using the idea of layering,the model gradually realizes the prediction of missing values,so as to improve the data density.Then,the time factors affecting recommendation are regarded as nodes in heterogeneous information networks,and a three-dimensional tensor model is constructed.Using the method of tensor decomposition,the time attributes are accurately and effectively integrated into the recommendation system.Experimental results show that the algorithm effectively improves the sparsity problem in the recommendation system.At the same time,the method of fusing influencing factors in the form of network entities can effectively improve the quality of recommendation.(2)Aiming at the problem of fusing preference information and non-scoring data,a collaborative ranking recommendation algorithm based on heterogeneous network graphs is proposed.The framework uses a variety of non-scoring data such as user preference relationship data,behavior data,etc.To construct a new heterogeneous information network relationship graph,which can intuitively show the potential relationship between users,items,and paired preferences.Then,using the generated heterogeneous network graph,a new sorting algorithm is proposed,which uses the meta-path in the heterogeneous information relation network to sort the recommendation results.Experiments show that compared with other benchmark algorithms,the algorithm improves the accuracy of recommendation and effectively improves the problems of data sparsity in recommendation system and cold start in collaborative filtering.(3)Aiming at the problems of multi-type heterogeneous network data fusion and network entity representation,a personalized recommendation method based on heterogeneous information network entity relationship representation is proposed.The algorithm focuses on fusing multiple types of network data to portray entities more vividly,and build a homogeneous entity relationship network based on this.Use network representation learning to obtain accurate entity representation vectors from homogeneous networks,and use the mapping matrix to fit the user and item representation vectors to predict the similarity score matrix.Experiments show that the fusion of multiple types of networks can help improve the quality of recommendations,and can maintain high accuracy under low sparsity.(4)Aiming at the problems of multi-source cross-domain recommendation and real-time recommendation in mobile applications,a multi-source real-time defect prediction recommendation algorithm in mobile applications is proposed,which is used for cross project defect prediction tasks in Android mobile applications and recommends the quality of code to developers in real time.The algorithm uses kernel based principal component analysis technology to transform the submitted instances into high-dimensional feature space to obtain representative features.Then use adversarial learning technology to extract the corresponding common feature embeddings.Experiments show that the predictive model can effectively use multi-source and cross-domain data to affect the quality of recommendation.Compared with other similar algorithms,it can and can obtain better results in various evaluation indexes.Focusing on the traditional problems in the recommendation system and the integration of various types of data,in order to improve the performance and quality of the recommendation system,this paper designs a method of using heterogeneous information network combined with recommendation model.The proposed method realizes the fusion of non scoring data,multi network types and cross domain data,improves the sparsity problem,and shows good recommendation performance in multiple experimental tests.