estim rate rare event multipl hierarchi scalabl log linear model deepak agarw rahul agraw rajiv khanna nagaraj kota problem estim rate rare event high dimension multivari categor data dimens hierarch problem routin data mine applic includ comput advertis main focu paper propos lmmh log linear model method scale massiv data applic billion train record million potenti predictor map reduc framework method exploit correl aggreg observ multipl resolut work multipl hierarchi stabl estim coarser resolut provid inform prior inform improv estim finer resolut predict accuraci scalabl method inbuilt variabl screen procedur base spike slab prior parsimoni remov inform predictor hurt predict accuraci perform larg scale experi data real comput advertis applic illustr approach dataset billion record hundr million predictor extens comparison benchmark method show signific improv predict accuraci 