larg scale behavior target ye chen dmitri pavlov john canni behavior target bt leverag histor user behavior select ad relev user displai state art bt deriv linear poisson regress model fine grain user behavior data predict click rate ctr user histori design implement highli scalabl effici solut bt hadoop mapreduc framework parallel algorithm result system build bt categori model entir yahoo user base dai scale imagin prior system approach yield ctr lift exist product system leverag ground probabilist model fit larger train dataset specif major contribut includ mapreduc statist learn algorithm implement achiev optim data parallel task parallel load balanc spite typic skew distribut domain data place featur vector gener algorithm linear time complex granular slide target window memori cach scheme significantli reduc number disk io make larg scale learn practic highli effici data structur spars represent model data enabl fast model updat work make signific contribut solv larg scale machin learn problem industri relev gener final report comprehens experiment result industri proprietari codebas dataset 