Simulated Annealing Based Algorithm for Tuning LDA Hyper Parameters

LDA is a very popular unsupervised model used to find thematic information about the documents. The performance of the LDA greatly depends on its hyper parameter values. If its parameters are tuned properly then LDA may produce much better results. This paper mainly focused on finding good LDA configurations for finding improved LDA output as compare to the traditional LDA model. We have proposed and implemented SA-LDA algorithm that uses Simulated Annealing (SA) to find optimal values of LDA parameters. An empirical evaluation using customer review datasets from three different domains namely Mobile, Hotel and Movie is conducted. The experiment results show that SA-LDA gives better performance when it is evaluated by the coherence score.