无界意义下在线变化分位数回归算法.pdf

数 学 杂 志 Vo1．34(2014) NO．2 J．ofMath．(PRC) VARYING QUANTILEREGRESSION WITH ONLINE SCHEM E AND UNBoUNDED SAM PLING WANG Bao—bin ．YIN Hong。 (J．SchoolofMathematicsandStatistics，CentralSouthUniversityforNationalities， Wuhan43oo74，China) (2．School Inform。ti。n，RenminUniversityofChin。，Bering100872,China) Abstract：W elconsiderakernel-based onlinequantileregression algorithm associatedwith asequenceofinsensitivepinballlossfunctions．Byiterationmethod andcomparisontheorem ，we obtain theerrorboundbased onthemoregeneraloutputspace． Keywords： quantileregression；Pinballloss；reproducingkernelHilbertspace；onlinealgo— rithm 2010MR SubjectClassiflcation： 60G15；62H05 DocumentCOde- A ArticleID： 0255．7797f2014)02．0281．06 1 Introduction W econsiderakernel--basedonlinequantileregressionproblem whenthesamplingpro-- cessisunbounded． LetX bean inputspace．y = R bean outputspaceand denote Z=X ×Y．Learningalgorithmisbasedonsamplesz= {(，y,)ILi1∈ZT，Tisthe samplesize，whicharedrownindependentlyfrom theBorelmeasureP．Inpreviousresearch， a lotofworkwasdonewhen the learning schemeisinvolved with theleastsquare loss ()= ，∈R．Thecorrespondingtargetfunctionistheregressionfunction ：X__÷y by 厂 x)= ／ydp()，∈X， wherepx()·istheconditionaldistributionofPateachX∈X．However，algorithmswiththe leastsquarewillloserobustnessifthedistributionofthenoisehasheavytailorabnormal variance．In 1964，Huberproposedtheleastmodulusmethod insteadoftheleastsquare， whichweakenedthenoisecondition．Forsparsity，Vapnik[5]combinedtheleastmodulus methodwiththethresholdvalueandintroducedtheg—insensitiveloss u)=(I钆I一￡)+， ∈ Received date：2013—09—22 Accepted da