Improved Evolutionary Approach for Tuning Topic Models with Additive Regularization.

The paper addresses a problem of tuning topic models with additive regularization by introducing a novel hybrid evolutionary approach that combines Genetic and Nelder-Mead algorithms to generate domain-specific topic models with better quality. Introducing Nelder-Mead into the Genetic Algorithm pursues the goal of enhancing exploitation capabilities of the resulting hybrid algorithm with improved local search. The conducted experimental study performed on several datasets on Russian and English languages shows noticeable increase in quality of the obtained topic models. Moreover, the experiments demonstrate that the proposed modification also improves the convergence dynamics of the tuning procedure, leading to a stable increases in quality from generation to generation.