Interest Topic Analyses And Location Prediction Of Users In Online Social Networks

With the development of society, there has been an increasingly obvious phenomenon of audience segmentation during people's information propagation. So user classification has become an important research topic. In the social network, if we can accurately classify the users, find similar online social networks and predict geographical locations of the audience, then this will be greatly helpful in understanding user characteristics, enhancing information propagation effectiveness, and improving user experience. Therefore, this paper is focused on classification of subjects of users'interests, multi-property detection of similar users, and prediction of geographical locations of video audience. Details are as follows:(1) Three methods for classifying microblog users are proposed. ?The information content-based method for classifying microblog users is proposed. First, the LDA subject model is used to extract topic distribution of each user as the feature value. Then, multiple models, such as the support vector machine, is used to classify users. ?The method that classifies users based on follower topology is proposed. Based on the observation that the users who share the same interest tend to have the same followers, the averaging method is used to calculate the intersection sets of followers as the feature value of each user. Then, multiple models, such as the support vector machine are used for user classification. ?Two combining methods for comparing probability estimations and merging feature values are proposed. The two proposed combining methods can combine the results from the information content-based approach and the follower topology-based approach in order to improve classification accuracy.(2) The method that can find similar users in term of multiple properties based on intuitionistic fuzzy sets is proposed. First, the difference of a particular property between two users is used to compute the similarity degree and difference degree, constructing an intuitionistic fuzzy number representing the similarity degree in this property. Then the aggregation operator is used to compute an intuitionistic fuzzy number as the integrated multiple-property similarity degree via the intuitionistic fuzzy number of each property. Finally, the score and accuracy of each integrated intuitionistic fuzzy number are computed and sorted. Users who are similar in multiple properties are those who are high on the list of score and accuracy.(3) The method for predicting the location of online videos' audience based on K-neighbor multi-label classification is proposed. First, the audience location prediction problem is converted into the multi-label classification problem. Predicting the location of online video's audience is equivalent to predicting the ranking of regions where videos are popular. If the online video is regarded as the sample and the audience's region is regarded as the label, then the problem of predicting the location of the audience is converted to the multi-label classification problem. Then, the classic K-neighbor multi-label classification method (ML-KNN) is improved in two ways. ? The weight-based method for measuring the similarity among samples is introduced and the method for computing feature weights is proposed. ?The algorithm for quickly finding similar samples is proposed. Based on these, the K-neighbor multi-label classification-based method for accurately predicting locations of online videos' audience (AL-KNN) is proposed.Based on a great deal of Twitter data, the performances of classifying users and finding similar users are evaluated. And based on a large volume of YouTube data, the experiments of predicting audience location are conducted.