Research On Recommender Algorithms For Data Sparsity Problem

Recommender system is an effective tool to solve the problem of information overload,and it is also a hot research field which is widely concerned by the academic and industrial circle.However,in practical applications,Recommender System is facing many challenges,such as the problem of data sparsity,scalability,cold-start,and interpretation,in which the problem of data sparsity is an important factor restricting the effect of recommendation.Aiming at the problem of data sparsity,this thesis focuses on the rating-sparsity problem in collaborative filtering recommendation,the trust-sparsity problem in social recommendation and the problem of data sparsity in context-aware recommendation,and improves several related recommender algorithms.The main work and innovation are as follows:1)In view of the rating-sparsity problem in collaborative filtering recommendation,this thesis defines Similarity Reliability,and proposes SSRCF(Similarity & Similarity Reliability-based Collaborative Filtering recommendation)algorithm.Firstly,the similarity reliability computing method is designed based on common rating numbers.Then,based on the Probabilistic Matrix Factorization model,the algorithm constrains the cosine of latent feature vectors by similarity relations,and controls constraint weight by similarity reliability.Due to the restriction of similarity reliability,the algorithm effectively reduces the impact of the rating-sparsity problem on similarity relations and strengthens the learning ability of the model for important similarity relations.which results in alleviating the rating-sparsity problem.The experimental results show that the proposed algorithm is obviously better than other current methods for collaborative filtering recommendation.2)In view of the trust-sparsity problem in social recommendation,this thesis proposes the concept of Social Liveness,and designs SocialST(Social Recommendation based on Social Liveness and Trust Enhancement)algorithm.Firstly,based on the PageRank algorithm,the LivenessRank algorithm to compute the social liveness is designed,which is used to treat the impact of social relations on users according to different weights,thus alleviating the trust-sparsity problem and improving the effect of recommendation.Then,when modeling social relation,the algorithm proposes a simple and efficient trust enhancement model based on SocialMF algorithm,simulating the relationship between trust intensity and friend preference similarity more accurately through power growth relationship,and further improving the effect of recommendation.The experimental results show that the proposed algorithm is obviously better than other current methods for social recommendation.3)In view of the problem of data sparsity caused by contextual information in context-aware recommendation,CARICC(Context-Aware Recommendation based on Item-grain Context Clustering)is proposed in this thesis.Firstly,based on the K-means algorithm,ICC(Item-grain Context Clustering)algorithm is proposed for mining high quality context cluster information.Then,the context information is modeled from the perspective of context clusters,and is integrated in the low-dimensional context clustering factor.Compared with complex and diverse contextual information,the lowdimensional context clustering factor carries rich knowledge,resulting in large-scale reduction of model parameters,thus alleviating the problem of data sparsity in contextaware recommendation effectively.The experimental results show that the proposed algorithm is obviously better than other current methods for context-aware recommendation.