Semi-Supervised Collective Matrix Factorization for Topic Detection and Document Clustering

Abstract

Topic detection and tracking (TDT) under modern media circumstances has been dramatically innovated with the ever-changing social network and inconspicuous connections among participants in the internet communities. Apart from the inherent word features of analysing materials, such as news articles and personal or professional comments, incidental information attracts increasing attention from the research community. Meanwhile, numerous interrelations hiding in the propagated articles and network participants also promote the transfer and evolvement of topics, not only apparent connections, for example having the same tags and belonging to the same party, but also weak connections which are complicated and with little causal relations. Therefore, answering the question how to exploit and use this hidden information in the social network will extend the landscape of research on TDT. In this paper, we employ the followers' groups extracted from Twitter as the social context that accompanied the corresponding news articles and explore the interior links among them to develop a non-negative factorization methods with semi-supervised information derived from the original data. Furthermore, experiments are conducted on real and semi-synthetic data sets to test the performance of topic detection and documents clustering. The results demonstrate that the proposed method outperforms several state-of-the-art methods.