discrimin topic model base manifold learn seungil huh stephen fienberg topic model popularli data analysi domain includ text document previou topic model probabilist latent semant analysi plsa latent dirichlet alloc lda shown impress success discov low rank hidden structur model text document model account manifold structur data gener inform linear dimension reduct map recent model laplacian plsi lapplsi local consist topic model ltm incorpor local manifold structur topic model shown result benefit approach fall short full discrimin power manifold learn enhanc proxim low rank represent neighbor pair consider neighbor pair paper propos discrimin topic model dtm separ neighbor pair addit bring neighbor pair closer preserv global manifold structur improv local consist present model fit algorithm base gener em concept pareto improv result dtm achiev higher classif perform semi supervis set effect expos manifold structur data provid empir evid text corpora demonstr success dtm term classif accuraci robust paramet compar state art techniqu 