Font Size: a A A

Research On Key Technology Of Social Network Recommendation System Based On Collaborative Filtering

Posted on:2016-05-11Degree:DoctorType:Dissertation
Country:ChinaCandidate:D H ChengFull Text:PDF
GTID:1108330473456378Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Nowadays, information is overload and users usually cannot make proper decision when they’re in the face of many choices. Now, personalized recommendation system builds a bridge from the user to the item, filtering irrelevant and uninteresting information. It presents users of their preferences items and solves the problem of information overload. Collaborative filtering algorithm is the most widely used technology of personalized recommendation. However, with the continuous development of online social network Social network, due to the complexity social network and the sparsity of user feedback data, the quality of recommendation in social network environment of traditional collaborative filtering algorithm is not good enough.In view of the above problems, this thesis focus on the effect of social relations, user cognitive level and user influence on User influence recommendation algorithm, and then put forward a new collaborative filtering recommendation based on social network social recommendation algorithm. Based on the algorithm, the social network recommendation system prototype is proposed in this thesis. The main research and results of the thesis are as follows:1. Due to the sparsity of user data, traditional similarity measurement algorithm cannot measure the similarity between users accurately. A similarity measurement algorithm using rating difference information entropy is proposed. The interest similarity between users is calculated by the modified function of information entropy. Not only the rating difference between users is considered in the modified function of information entropy, but also the user’s rating expectation and the number of common rating. And then, the final similarity is calculated as the weighted sum of interest similarity and social familiarity. The experimental results on real data demonstrate that the proposed similarity algorithm alleviates the influence of the sparsity of data to a certain extent and enhances the recommendation quality.2. According to the statistical law of real data and the analysis on the experimental results of trust-based recommendation, the quality of recommendation is strongly affected by the user cognitive level. When the user cognitive level is low, adding the trust relationship is helpful for the quality of recommendation which is inversely proportional to the distance of propagation. When the user cognitive level is high, the quality of recommendation is not sensitive to the trust relationship and increasing the propagation distance can significantly improve the recommendation coverage. By the mentioned above, a kind of classification recommendation algorithm based on the user cognitive level is proposed. According to the difference of the user cognitive level, the weights of the trust relationship and the interest similarity are different. The experimental results show that compared with the traditional trust-based recommendation algorithm, the proposed algorithm obtains better prediction performance, as well as good scalability.3. There’re two significant problems of social network analysis, the importance of the individual and the relationship between users in the social network, investigated in this thesis profoundly, and integrated into the social recommendation based on matrix factorization. Based on user influence and adaptive similarity, an algorithm of matrix decomposition is proposed in this thesis. ClusterRank, a semi local algorithm, is used to measure the influence of users, which not only takes the users’ own influence into consideration, but also the effect of clustering coefficient on information dissemination. In this algorithm, the similarity of users is obtained by the matrix decomposition in the iteration learning process, as well as the user feature vector, which solves the problem of similarity cannot be directly measured without the common rating items between users.4. Based on the work of the previous chapters, the prototype design of social recommendation system is proposed in this thesis. Not only the function realization of the traditional recommendation system is considered in this design, but also the social network information (circle of friends) is integrated into the prototype design of recommendation system. The module of the circle of friends can provide the trust relationship, the measurement information of users, and the similarity between users to the recommendation algorithm. Meanwhile, it also can adopt the recommendation by recommendation algorithm to further enrich relationship of users.By the analysis of user interest, user rating law and social network, various key technologies of social recommendation is researched deeply in this thesis, such as the similarity between users, the trust relationship between users, user cognitive level, user influence, and the model of user interest in the social network is established, and reliable personalized recommendation algorithm is proposed, as well as the prototype design of social recommendation system, which providing valuable references for the further realization of high-quality personalized recommendation to users in social network.
Keywords/Search Tags:Social Network, Collaborative filtering, Similarity Measurement, User Cognitive Level, Trust-based Recommendation, Matrix factorization, User Influence
PDF Full Text Request
Related items