|
|
自动化学报 2008
先验知识与基于核函数的回归方法的融合DOI: 10.3724/SP.J.1004.2008.01515, PP. 1515-1521 Keywords: Machinelearning,priorknowledge,kernelbasedregression,iterativegreedyalgorithm,weightedlossfunction Abstract: ？Insomesamplebasedregressiontasks,theobservedsamplesarequitefewornotinformativeenough.Asaresult,theconflictbetweenthenumberofsamplesandthemodelcomplexityemerges,andtheregressionmethodwillconfrontthedilemmawhethertochooseacomplexmodelornot.Incorporatingthepriorknowledgeisapotentialsolutionforthisdilemma.Inthispaper,asortofthepriorknowledgeisinvestigatedandanovelmethodtoincorporateitintothekernelbasedregressionschemeisproposed.Theproposedpriorknowledgebasedkernelregression(PKBKR)methodincludestwosubproblems:representingthepriorknowledgeinthefunctionspace,andcombiningthisrepresentationandthetrainingsamplestoobtaintheregressionfunction.Agreedyalgorithmfortherepresentingstepandaweightedlossfunctionfortheincorporationstepareproposed.Finally,experimentsareperformedtovalidatetheproposedPKBKRmethod,whereintheresultsshowthattheproposedmethodcanachieverelativelyhighregressionperformancewithappropriatemodelcomplexity,especiallywhenthenumberofsamplesissmallortheobservationnoiseislarge.
|
