fast onlin learn offlin initi time sensit recommend deepak agarw bee chung chen pradheep elango recommend problem larg dynam item pool ubiquit web applic content optim onlin advertis web search avail rich item meta data excess heterogen item level requir inclus item specif factor weight model estim item factor computation intens pose challeng time sensit recommend problem import rapidli learn factor item new articl event updat tweet onlin fashion paper propos method call fobfm fast onlin bilinear factor model learn item specif factor quickli onlin regress onlin regress item perform independ procedur fast scalabl easili paralleliz converg independ regress slow due high dimension central idea approach larg amount histor data initi onlin model base offlin featur learn linear project effect reduc dimension estim rank linear project take recours onlin model select base optim predict likelihood extens experi show method significantli uniformli outperform competit method obtain rel lift rang term predict log likelihood rank correl metric proprietari yahoo dataset obtain reduct root squar error previous method benchmark movielen dataset time base train test data split 