Finding anchor words of separable-nonnegative matrix factorization based on singular value decomposition

Topic detection is a process to find topics or subjects of discussion in a collection of documents such as tweets on Twitter. Manual detection of topics on Twitter is difficult because of too many tweets. Therefore, it is necessary to detect topics automatically. One of the automatic methods for topic detection is the Separable-Nonnegative Matrix Factorization (SNMF) method with the AGM algorithm. SNMF is a matrix factorization-based model that can be solved directly using the assumption that each topic has one word, called anchor words, that is not present in other topics. SNMF with AGM algorithm consists of three stages, namely the constructing the co-occurrence matrix, finding the anchor words, and recovering the topics. The common method to find the anchor words is the convex hull-based method. In this paper, we examine the process of finding anchor words based on Singular Value Decomposition (SVD). The results show that by considering all words as anchor word candidates, the SVD-based method gives better results than the convex hull-based method. Meanwhile, when the anchor finding was done by using anchor threshold, the convex hull-based method still gives a better result than the SVD-based method.