effici method topic model infer stream document collect limin yao david mimno andrew mccallum topic model provid power tool analyz larg text collect repres high dimension data low dimension subspac fit topic model set train document requir approxim infer techniqu computation expens todai larg scale constantli expand document collect infer topic distribut document retrain model paper empir evalu perform method topic infer previous unseen document includ method base gibb sampl variat infer method inspir text classif classif base infer method produc result similar iter infer method requir singl matrix multipl addit infer method present sparselda algorithm data structur evalu gibb sampl distribut empir result sparselda approxim time faster tradit lda provid speedup previous publish fast sampl method substanti memori 