Research On Collaborative Filtering Recommendation Algorithms Based On Social Relationships And Communities

With the fast development of Internet,the information overload problem is becom-ing more and more serious nowadays.Recommender Systems(RSs)are proposed to tackle this problem.Generally,recommendation algorithms are developed to analyze historical behavior data of users,to mine their personal interests,as well as to recommend the infor-mation and services which users may be interested in.RSs can help filter useful information for users which definitely saves our time and improves the utilization ratio of information resources.Currently,RSs have been employed in E-commerce websites,Video websites,Music websites and some other personalized advertisements services,which have produced huge economical benefits.Currently,Collaborative Filtering(CF)algorithms are the most widely investigated and implemented recommendation techniques.CF algorithms rely on the "user-item" ratings information to model user preferences,and predict the future interests of users.However,traditional CF algorithms have some inherent weaknesses,such as the data sparsity problem,cold start problem,efficiency and scalability problem etc.In the past few years,some works have been done to incorporate users' social relationships into the traditional CF algorithms,and try to introduce the community mining techniques into CF algorithms.The results of existing works show that the data sparsity problem and cold start problem is effectively alleviated by incorporating the social relationships,in the meantime,the employment of community can alleviate the scalability problem to a large extent.Therefore,the research of collaborative filtering algorithms based on social relationships and community is very meaningful.The main contributions of this dissertation are summarized as follows:(1)Existing works about social relationship based collaborative recommendation al-gorithms usually assume that users are similar with their friends,regardless of the hetero-geneity and diversity of social relations.In order to tackle this problem,we first formally define a user-interest preference matrix to describe the fine-grained interests of users,and then employ a fuzzy clustering method to partition users with multi-aspect similar inter-ests into the same clusters and generate the fine-grained friends subsets.By proposing a clustering-based social regularization term,we incorporate the fine-grained social relation-ships into the traditional matrix factorization model,which confines the user feature vectors with the feature vectors of find-grained similar neighbors.Real world datasets based exper-imental results demonstrate that our method outperforms the state-of-art method in terms of prediction accuracy.(2)Existing social relationship based algorithms mainly consider the explicit social relationships,and ignore the implicit social relationships.However,in the real world,the explicit social relationships are generally very sparse or unavailable.In the meantime,some explicit relationships are not in high-quality.If the recommendation algorithms rely on these low-quality social relationships,the prediction accuracy will definitely be decreased.In or-der to tackle this issue,in this paper,we take trust relationships as an example,and propose an implicit trust relationship prediction algorithm.By analyzing the important factors which influence the establishment of trust between users,the algorithm can discover the latent rela-tionships between users in trust networks automatically.Then we incorporate the predicted implicit trust relationships into some trust based recommendation algorithms.Real world datasets based experimental results show the effectiveness of the implicit trust mining al-gorithm.The comparison experimental results demonstrate the effectiveness of employing implicit trust into three representative social recommender systems.In conclusion,the pre-dicted implicit social relationships are capable of assisting or replacing the explicit trust relationships when they are very sparse.(3)The main idea of community based recommendation algorithm is to mine user or item communities based on users' historical rating information by employing community mining techniques,and then perform the traditional recommendation algorithms within each community.The community based recommendation methods are effective in tackling large scale data sets by reducing the dimension of data.However,since the prediction can just be made within certain community,the decrease of data will result in sparser similar neigh-bors or less similar neighbors for the memory-based recommendation methods compared to those discovered in the whole dataset.Therefore,both the prediction accuracy and coverage are comparatively low.To tackle this issue,some methods explore to mine communities based on additional information,e.g.,trust relationships,which present a new way to find similar neighbors.However,existing trust community based recommendation methods gen-erally ignores the distrust information in trust networks,as well as the serious data sparsity problem existing in trust communities.In order to solve these problems,we propose a rec-ommendation algorithm which combines a trust community mining algorithm and a sparse rating filing algorithm.In order to cluster the trust and distrust relationships simultaneously,we propose a SVD signs based trust community mining method,which divides the users connected by trust relationships into the same trust community and the users in different communities distrust each others.In addition,we propose a missing rating filling algorithm to alleviate the sparsity problem.By mining the ratings of the trust neighbors in the same trust community,the missing ratings of the target users are complemented to some extent.Real world datasets based experimental results shows that the trust communities generated by the clustering of trust and distrust relationships are meaningful.On the other hand,the missing rating filling algorithm is able to alleviate the data sparsity problem effectively,as well as to improve the prediction accuracy and coverage.(4)In the model-based recommendation algorithms,less data can be used for model training after the community partitioning compared to the original dataset,which will influence the improvement of recommendation accuracy.Since the performance of community based recommendation algorithms is sensitive to the communities,and different community structures generally show different data correlations.Therefore,we propose a novel community based collaborative filtering recommendation method which systematically considers different community structures,e.g.,the similarity based user community,the trust based user community and the similarity based item community.In this paper,we investigate how to incorporate multiple community structures into a unified recommendation framework,and explore the influence of different community structure to the recommendation accuracy.Real world datasets based experimental results demonstrate that the proposed approach outperforms the state-of-the-art algorithms in terms of prediction accuracy.What's more,we have discussed the superiority of the proposed algorithm in tackling data sparsity,cold start and scalability problems.