fast collaps gibb sampl latent dirichlet alloc ian porteou david newman alexand ihler arthur asuncion padhraic smyth max well paper introduc collaps gibb sampl method wide latent dirichlet alloc lda model method result signific speedup real world text corpora convent gibb sampl scheme lda requir oper sampl number topic model propos method draw equival sampl requir averag significantli oper sampl real word corpora fastlda time faster standard collaps gibb sampler lda approxim show fast sampl scheme produc result standard slower sampl scheme experi real world data set demonstr speedup wide rang collect size pubm collect million document requir comput time cpu month lda speedup save cpu month comput 