Matrix Inter-joint Factorization - A New Approach for Topic Derivation in Twitter

Amongst all the social media platforms available, Twitter is rapidly becoming the main one used for communications about real-time events. As a result, there is a lot of interest in monitoring Twitter and understanding the topics of conversations. However, the fact that tweets are short in content makes topics derivation a challenge, as most existing methods use content features only, sometimes integrated with limited interaction information. In this paper, we propose a novel method: Non-negative Matrix inter-joint Factorization (NMijF), in which we conduct co-factorization jointly over Twitter interaction features and content attributes within a single iterative-update process. We have conducted comprehensive experiments on real Twitter datasets and evaluated the performance of the proposed method, especially comparing it with the Joint Non-negative Matrix Factorization (joint-NMF) and Non-negative Matrix co-Factorization (NMcF) methods. Our experiment results show that the proposed NMijF method outperforms joint-NMF, NMcF and other advanced topic derivation methods in terms of Topic Coherence, Purity, Normalized Mutual Information and Precision-Recall.