Research On Multi-label Classification In Social Network Environments

With the rapid development of social network, lots of social network sites with massive user have emerged, such Facebook, Twitter, and YouTube. As a medium to share knowledge and connect with friends, social networks are playing an increasingly important role in our lives. Label classification is an important application in social networks. For instance, users have interest labels and friend relation labels in social networks. Besides, they can also tag various texts, images and videos in social networks. In traditional label classification, networked data is represented by a single label. But with the enrichment of social network applications, the form of networked data is increasingly diverse, and single label cannot effectively reflect the complex and multi-semantic features of social network data. Therefore, the research of multi-label classification in social network environment has got more and more attention. Based on this, our paper mainly focuses on three aspects, which are social network structural analysis, multi-label classification in social networks and the application of multi-label on recommendation systems. The main work is as follows:(1) We introduced the background and significance of multi-label classification in social network environments. Analyzed the research status and drawback of social network structural analysis, multi-label classification, and recommendation system areas, and demonstrated the concept, classification, key parameters and classical algorithms of the three areas.(2) We proposed a link life based social network structural analysis method. We add link life into social network structural analysis to test the influence of link life on some important basic parameters of social networks (including degree distribution, network diameter and average clustering coefficient). Experiments showed that, after adding link life, the evolutions of social network structure are different with results of traditional research. Specifically, a small change in link life will cause dramatic changes in network diameter.(3) Based on the above social network structure, we proposed two semi-supervised multi-label classification algorithms. On the basis of two classical relational classifiers, we add must-link constraint and uncertainty probability to evaluate the influence of must-link constraint on multi-label classification. Experiments showed that our methods have got better accuracy and efficiency than traditional relational classifiers on large scale social networks, especially when known labels are few.(4) Based on the social contexts, which were calculated by using the above algorithms, we proposed a Multi-source Rating Aggregation algorithm. First, we calculate the academic expertise of recommenders by using their social contexts. Then, we add academic expertise into multi-source rating aggregation process to get a more accurate recommendation. Experiments showed that our method can effectively eliminate the interference noise, which is brought by strict and lenient recommenders in recommendation systems, and do not need any prior information about the ratio of strict and lenient recommenders.