A multi-label model to predict undisclosed attributes in microblogging

In this paper, the automatic prediction of micro-blog users´ attribute information is modeled as a multi-label classification process considering the relevance and co-occurrence of user attributes instead of single-label or regression, where each user has more than one attribute. Two basic and one improved multi-label classification methods are evaluated and compared in this task. Taking micro-blogging as the background, we conduct a series of experiments, comparison and analysis to prove that the multi-label classification is a very effective method to predict the user profiles considering the relevance and co-occurrence of user attributes. In feature selection stage, we consider a variety of attribute information of users such as user nickname, user tags and user personal description. The results show that a relatively great performance is obtained. Finally, we analyze the user profiling and the user preferred vocabulary in the associated attributes.