inform theoret regular semi supervis boost lei zheng shaojun wang yan liu chi hoon lee present semi supervis boost algorithm increment build linear combin weak classifi gener function gradient descent label unlabel train data approach base extend inform regular framework boost bear loss function combin log loss label data inform theoret measur encod unlabel data inform theoret regular term make optim convex propos simpl sequenti gradient descent optim algorithm obtain impress improv result synthet benchmark real world task supervis boost algorithm label data state art semi supervis boost algorithm 