Font Size: a A A

Research On Recommendation Algorithm Based On Matrix Factorization

Posted on:2019-08-25Degree:DoctorType:Dissertation
Country:ChinaCandidate:S L ChenFull Text:PDF
GTID:1368330611993009Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Internet and information technology have changed the way people discover and obtain information,and have brought great convenience to people's work and study.With the rapid growth of the number of users,the information on the Internet is also expanding rapidly.An enormous amount of information can meet users' different needs for various information better,however,it also poses a serious challenge to information processing technology.Owing to the variety,uneven quality and complex structure of information resources,it is becoming more and more difficult to find valuable information from the vast flood of information.This is the so-called ”information overload” problem.In this context,recommendation systems emerge as the times require.Recommendation systems analyze users' preferences based on historical behavior data and recommend content that may be of interest to them.As important tools to solve the problem of information overload,recommendation systems have been widely used in many fields,such as electronic commerce,personalized website and so on.Despite the great success,recommendation systems still face some urgent problems that need to be solved,mainly including the ”sparsity” and ”cold start” problems.To address the above problems,this paper puts forward two research routes: One is to study how to make full use of the existing rating information to alleviate the sparsity and cold start problems;the other is to study how to solve the sparsity and cold start problems by integrating auxiliary information.In this paper,we deeply investigate the recommendation algorithms in rating prediction and Top-N recommendation scenarios.The main innovations can be summarized as follows.(1)We propose a cluster feature based latent factor model.In the rating prediction task,most existing recommendation algorithms only utilize users' or items' global or local information for prediction,but fail to make full use of the information contained in the rating data.Aimed at this problem,the paper describes a rating prediction model by incorporating local and global information.Specifically,we first extract users' and items' local clustering features by fuzzy clustering method.Then by incorporating this kind of local information,we design an integrated latent factor model.Compared with the traditional approaches,our approach can make full use of the rating data and get higher prediction accuracy.(2)We propose a semi-supervised model for Top-N recommendation.In the TopN recommendation task,there are also sparsity and cold start problems.Most Top-N recommendation algorithms ignore users' preference among the unrated items and can not achieve satisfactory recommendation performance on sparse data sets.Aimed at this problem,the paper presents a Top-N recommendation model based on semi-supervised learning.Unlike the traditional algorithms,which only consider users' preference between the rated items and the unrated items,our model takes the rated items of each user as a positive sample set,and further divides the unrated items of each user into two subset: an intermediate set and a negative sample set.Then,we make full use of users' preference between the three item sets.In addition,we propose a semi-supervised learning method to divide users' unrated items.The experimental results on several real data sets demonstrate our approach significantly outperforms traditional Top-N recommendation models for all evaluation metrics.(3)We propose a matrix factorization model with explicit and implicit feedback.Numerous existing matrix factorization models rely heavily on explicit feedback.When the explicit feedback is relatively sparse,these models always perform poorly.In order to address this problem,the paper designs a latent factor model based on probabilistic matrix factorization,by incorporating implicit feedback as complementary information.In the proposed model,the explicit and implicit feedback matrices are decomposed into a shared subspace simultaneously.Then,the latent factor vectors of users and items are jointly optimized.The experimental results using the MovieLens datasets demonstrate that the proposed algorithm outperforms the baselines on both rating prediction and TopN recommendation tasks.(4)We propose a matrix factorization model fusing review information.Users' reviews on items are important feedback information,from which we can extract users' preferences,items' characteristics and other valuable information.Making full use of the information in the reviews can help solve the sparse and cold start problems in the recommendation systems.Traditional recommendation algorithms usually focus on users' rating information but ignore the review information.Aimed at this issue,this paper proposes a matrix factorization model by fusing review information.Firstly,we extract the latent features of users and items from the review data by bag-of-word model or document embedding model.Then the extracted latent features are fused to matrix factorization model.Extensive experiments and analyses on Amazon data sets show that the model can effectively alleviate the sparsity and cold start problems.In summary,this paper studies the key issues of rating prediction and Top-N recommendation in recommendation systems.We make full use of the existing information and integrate auxiliary information to improve the performance of the recommendation algorithms,and verify the effectiveness of the proposed algorithms on real data sets.The research work in this paper has certain theoretical and practical significance for recommending system application.
Keywords/Search Tags:Recommender Systems, Recommendation Algorithms, Collaborative Filtering, Matrix Factorization
PDF Full Text Request
Related items