Comparative study of original recover and recover KL in separable non-negative matrix factorization for topic detection in Twitter

An increasing amount of information on social media such as Twitter requires an efficient way to find the topics so that the information can be well managed. One of an automated method for topic detection is separable non-negative matrix factorization (SNMF). SNMF assumes that each topic has at least one word that does not appear on other topics. This method uses the direct approach and gives polynomial-time complexity, while the previous methods are iterative approaches and have NP-hard complexity. There are three steps of SNMF algorithm, i.e. constructing word co-occurrences, finding anchor words, and recovering topics. In this paper, we examine two topic recover methods, namely original recover that is using algebraic manipulation and recover KL that using probability approach with Kullback-Leibler divergence. Our simulations show that recover KL provides better accuracies in term of topic recall than original recover.