Research On Recommendation Algorithm Based On Heterogeneous Information Network Representation Learning

The main aim of recommender system is to actively recommend information and commodity which matches users' preference based on past user item interaction information and auxiliary information.The recommender systems are widely deployed in modern E-commerce systems,and the performance of recommender system will significantly affect the revenue of merchants and user experience.However,most recommendation algorithms are apt to suffer from cold start and data sparsity problem.In order to alleviate this problem,the common used method is to utilize auxiliary information to enhance recommendation performance.In recent years,because recommender system contains different types of nodes and edges,using Heterogeneous Information Network to model recommendation system has gained increasing attention.But most of the algorithms use path-based similarity to enhance recommendation,which can't fully utilize the high order proximity in HIN.Besides,the algorithm can't effectively combine the similarity from different meta paths.What's more,most local low rank matrix approximation algorithms randomly select anchor points,which can't effectively use abundant semantics in HIN.How to effectively utilize HIN to divide submatrix in Local Low-Rank Matrix Approximation is worth studying.For the problem of LLORMA and Collaborative Filtering stated above,this paper has mainly done the research work stated below:1.An algorithm HRLRec was proposed which can combine the collaborative filtering and Heterogeneous Information Network Representation Learning.Besides,HRLRec can learn user and item latent factor from rating records and combine embedding from different meta paths semantics to enhance recommendation,and use PathSim as a regularization term to enhance the latent representation of user and item.2.In order to divide local low-rank submatrix more effectively,this paper introduced embeddings learned from heterogeneous information network representation learning todivide the submatrix,and proposed an algorithm named HLORMA.This paper first learns high order proximity embedding from HIN,and then use different clustering methods to divide user and item into clusters.Then the algorithm use cluster centers as anchor points,and use kernel function and distance threshold to decide the elements of the submatrix.Afterwards,this paper utilizes matrix factorization in each submatrix and use weighted ensemble to predict the ratings,which can achieve performance gain compared to other methods.3.The proposed model and algorithm are implemented in Python,and are tested with real-world dataset in experiment section to compare with other algorithms to prove the effectiveness of the proposed method.Empirical result shows that our algorithm has the effective performance.The hyperpapameter settings used in the model are discussed.