Retrofitting Decision Tree Classifiers Using Kernel Density Estimation

A ]IOVC1 mdl)d for cxnnbining dccisio]l trcws a]ld kcn)d dmlsity cstlimators i s ]woposcd. Sta]ldard classification]) tmcs, or class prob al)ility trms, ]movidc piuwwisc constant estimates of class posterior ]mobabilitlics. KcrI)C1 dmlsity estimators can ])rovidc smooth ]Io]l-])alalllct]ic estimates of class probaliitics, l)ut scale ])oorly as the dilncl]sionality o f tllc ])roblcm illcrcascs. ‘J’his palm dismsscs a l)ylnid sclIcIIlc whicl] uscx decision trees to find t,llc rclcwal)t s t ructure ill IIig]l-(li]llc]lsio]lal classification lnoblmns and tllml uses local km)cl density esti]nat,cs to fit Slnootl) ]mobability cstilllates within t h i s structure. IJxlmrilnclltal mmlts 011 simulated da ta indicate that tl)c mctllod ~wovidcs sub stalltial i]nlmwmncl]t over trees or clmlsity ]I]etllods alo]lc for ccrtaill c lasses of ]moblmns. ‘1’llc I)al)cr briefly discusses various cxtcllsiolls of tllc basic a])lnoacll and t,llc types of a]q)]icatiou for wllicll tllc ]nctllod is Lest suitd.