Social Stream Classification with Emerging New Labels

As an important research topic with well-recognized practical values, classification of social streams has been identified with increasing popularity with social data, such as the tweet stream generated by Twitter users in chronological order. A salient, and perhaps also the most interesting, feature of such user-generated content is its never-failing novelty, which, unfortunately, would challenge most traditional pre-trained classification models as they are built based on fixed label set and would therefore fail to identify new labels as they emerge. In this paper, we study the problem of classification of social streams with emerging new labels, and propose a novel ensemble framework, integrating an instance-based learner and a label-based learner by completely-random trees. The proposed framework can not only classify known labels in the multi-label scenario, but also detect emerging new labels and update itself in the data stream. Extensive experiments on real-world stream data set from Weibo, a Chinese micro-blogging platform, demonstrate the superiority of our approach over the state-of-the-art methods.