Social Recommender System Based On Sentiment Similarity

With the development of Web2.0, social media has become the main platform for information dissemination, information access and communication. Social media bring more convenience today. But quick information update and information posted result in huge information redundant to communication platform. In order to exactly find out information and users that a user preferred form huge information, the help of intelligent system is need. Social recommendation system is an important tool to deal with information overload in social media platform. It can help users to find their interested users and content from the mass information. Research on social recommender systems can expand recommended range and accuracy in the user social network to complete knowledge active recommendation. It also can help enterprises to enrich the means of marketing and business growth. However, it is now a research challenge that how to use social network and text information to build the framewoek of social recommender system on social media. Therefore, in this paper, we take Sina Weibo platform as a case and study how to use all kinds of information to construct social recommendation system. The main research contents are as follows:Firstly, we study the data collection of micro-blog network based on variable precision and basicly analyze its network. When collecting data of micro-blog network for recommender system, we need to difine the network boundary. In this paper, we use variable precision to detemine the scope of collected data. We firstly test this method on public data set. Variable precision is used to handle movie rating data, and study changes of recommendation results in different accuracies. The comparison results show that the better precision of data, the higher accuracy of the recommendation only in a certain range of threshold. Therefore, it is verified that variable precision can determin the optimum boundary of collected data. Then we use this method to determin the seeds of micro-blog users. Using the recommendation method that used on movie rating recommendation, we study the recommendation results of different network data produced by variable precision. According to the results, we choose the optimum number of user seeds to determine the optimum data set for recommender system. Based on this, we collect micro-blog network data on a certain topic. Moreover, we basically analyze this network and check its character of scale-free.Secondly, we study the nodes link prediction based on exponential random graph mode for micro-blog social network. In this work, we propose exponential random graph model to model the diabetic micro-blog network, and use the Monte Carlo maximum likelihood estimates (Monte Carlo MLE) to calculate the parameters in different structures. Through simulation and fitting analysis, we compare the ability that exponential random graph model is constructed at what extent to represent the original networks. We model the diabetes microblog network as empirical study and illustrate the evolution of different model structure in different periods. At the end, we get the representive model. Since non-linked node get linked will cause configurations changed, we can computer the link probability of node pair by calculating gap between non-linked and after linked changes. The resort list can be used as recommdation items for social recommender system.Thirdly, we use opinion mining to study text features of micro-blog social network. Sentiment analysis (Opinion mining) is the attitude classification for text content and search user attitude from the text information and other information that user released on a certain topic. In micro-blog platform, information contains both text content and emoticons. Therefore, we select126text and micro-blog features on subject of diabetic in micro-blog. Fusion of information gain and support vector machine is used to extract useful feature later. And we classify these features for sentiment classification, and analyze features contribution of different categories. We calculate the average distance of positive and negative text based method of Karhunen-Loeve transform. Then the concept of sentiment similarity is proposed in this paper to measure the attitude correlations between users on the same topic.Finally, social recommendation system framework based on amended sentiment similar is constructed. We analyze the impact of three social psychological and behavioral factors, which are trust, homogeneity and opinion leaders. Those three factors influence the user when making a decision to take recommendation items. To the end of mapping those three social factors into micro-blog network, we quantify them with design formulas. Link prediction based on exponential random graph model and amended sentiment similar are bound together, which provide recommendation list conformed to user emotional preference. Thus we further build a social recommendation system framework based on above similarity measurement method. In order to validate the performance of this recommendation method, we empirical analyze user recommendation on the topics of diabetes and infants. Comparing with other methods, the evaluation results verify the superiority and general applicability of our method in social recommender system.