Research On Personalized Recommendation Algorithms In Social Network Environment

As social network gradually becomes universal, the amount of information, such as network commodities and rates are growing rapidly. Confronted with the massive data, users are unable to make an effective and efficient choice, which, therefore, pushes the generation of personalized search engines. The traditional recommendation algorithms only adopt users' historical rating information and provide two-dimensional recommendation, in real-world applications, they are subjected to many restrictions. Therefore, the recommendation quality turns out to be poor and low.In view of the above problems, this dissertation aims to probe into the ways to further improve the performance of social network recommendation from dimensional aspects. The adopted social network dimensional information includes: social network contextual information, social tag information, inter-user trust relations, etc.The main contents in the dissertation include the following four parts:(1) Research of Context-Aware Personalized RecommendationThe dissertation puts forward the contextual-based system for better constructing the model of user's preferences to further increase the performance of the recommendation system. The system adopts Radom Decision Tree with higher learning precision to integrate many kinds of contexts, which not only assures the lower complexity but also increases the accuracy of constructing the model of user's preference. Radom allocation strategy is adopted in the recommendation to categorize the original user-item rating matrix R. In this way, the rating of similar users or similar items can be divided into the same node in the decision tree, i.e. rating information with similar contexts are divided into the same group and the ratings in the same group enjoys higher relativity than those in the original rating matrix. The filtering of the context can increase the accuracy of establishing the model of user's preference so that the accuracy of the system is also intensified.(2)Research of Social Tag Recommendation Combining Topics with Linguistic ModelIn order to solove the problem that how to make use of the abundant tag information in social network to realize the social recommendation effectively, the dissertation takes the recommendation centered by users and resources into consideration to realize personlized social tag recommendation for users. This method combines that Language Model(LM) with Latent Dirichlet Allocation(LDA) to evaluate the accuracy of the new tags. The advantage of LDA's application lies in the generation of new tags which have never been used by users to increase the vocabulary applicable to users during tag recommendation and provide more accurate tag information when users mark tags on resources. The research from this part optimizes tag information dimension of users and items. It not only creates a new dimension to establish user preference model but also lay a foundation for the research on social network recommendation combining social tags with trust relations.(3)Research of Social Network Recommendation Based on Trust RelationsAs for the data sparsity and cold start problem, a new model combining user's rating with social trust relation called Trust-based Social Network Recommendation(TSNR) based on trust relations is proposed. This method takes negative effects of untrustworthy nodes impose on the trust-based social network recommendation into full consideration. Therefore, the first step of the algorithm is to spot the untrustworthy nodes in the trust network through computing the reputation and deviation value of every node and alleviate its negative effects on trust network by weakening its rating weight. Secondly, given that users' preferences can be influenced by their friends, it can revise the users' feature vector from their friends' trust matrix to solve the problems of users' feature vector accuracy establishment and trust transmission. And meanwhile, in order to minimize the round-off error, it realizes social network recommendation through matrix factorization with social regularization constraints. The results of experiments of TSNR on public dataset reveals that new recommendation has been greatly improved compared to the traditional collaborative recommendation especially when it is applied to the dataset with sparse or even missing user rating items, the recommendation still enjoys outstanding performances.(4) Research of Social Network Recommendation Combining Social Tags with Trust RelationsThe dissertation brings forward a social network recommendation algorithm combining social tags with trust relations called TTR, i.e.: Tag-based and Trust-based Recommendation. It collects major information relating to social trust relations, item tag information and user rating matrix based on probabilistic matrix factorization. All the data resources from different dimensions are connected through shared users' potential spaces(or item potential spaces). The above mentioned 2 types of spaces can be obtained by applying the probabilistic matrix factorization. In this way, effective social recommendation result can be achieved. The results generated from public datasets reveal that TTR is superior to the existing trust-based social recommendation or social tag recommendation algorithmsThe dissertation also conveys the innovative points as follows:1. To optimize user preference model by using contextual information and social tag information which provide a precise user model for latter socialized recommendation in order to improve the precision of social recommendation.2. To take the negative influence of incredible nodes on recommendation system into consideration and solve the equitable calculation problem of credibility based on the trust-based recommendation model.3. To properly evaluate the initial credibility with the help of credibility transmission mechanism in order to solve the new users' cool boot problem if new users have no information relating to rating data or social relations.