S7_extraFeatureSelection

download S7_extraFeatureSelection

of 7

Transcript of S7_extraFeatureSelection

  • 7/29/2019 S7_extraFeatureSelection

    1/7

    S7Extra:FeatureSelec/on

    ShawndraHill

    Spring2013

    TR1:30-3pmand3-4:30

  • 7/29/2019 S7_extraFeatureSelection

    2/7

    FeatureSelec/onStep1:

    UseDomainknowledgetoguideyouwheneverpossible

    Step2:

    VisualizeaKributes

    RemoveaKributeswithnovalues,toomanymissingvalues Checkforobviousoutliersandremovethem

    Step3:

    ConstructnewaKributes(ifitmakessense)

    CombineaKributes NormalizenumericaKributes(forregression,NaveBayes,NNhKp:www.tus.edu

    ~gdallalregtrans.htm)

    CreatebinaryaKributesfromnominalaKributesStep4:

    SelectthebestsubsetofaKributesfortheproblem

    IFINDOUBTCHOOSEAMETHODTHATDOESTHEFEATURESELECTIONFORYOU(forexample,

    decisiontrees)

  • 7/29/2019 S7_extraFeatureSelection

    3/7

    TheBasics

    BasicIdeasUsuallyfacedwithproblemofselec/ngsubsetofpossiblepredictors

    Havetobalanceconflic/ngobjec/ves Wanttoincludeallvariablesthathavelegi/matepredic/veskill Wanttoexcludeallextraneousvariablesthatfitonlysample-specificnoise

    Reducepredic/veskill Increasestandarderrorsofregressioncoefficients,classifica/on,etc.

    Ideallywouldbeabletodeterminesinglebestsubsetofpredictorstoinclude

    Butnosingledefini/onofbest Differentalgorithmswillproducedifferent"best"subsets Problemsmagnifiedbycorrela/onamongpredictors

  • 7/29/2019 S7_extraFeatureSelection

    4/7

    FeatureSelec/on

    RankingBysomeobjec/ve(forexample,informa/ongain)

    Subset

    Algorithms(seenextslide)Wrapper(trysubsetwithinthecontextofthealgorithmyouknowyouaregoingtouse)

  • 7/29/2019 S7_extraFeatureSelection

    5/7

    FeatureSelec/onAlgorithms

    Allpossiblesubsets Onlyfeasiblewithsmallnumberofpoten/alpredictors(maybe10orless) Thencanuseoneormoreofpossiblenumericalcriteriatofindoverallbest

    Forwardstepwiseregression Startwithnopredictors

    Firstincludepredictorwithhighestcorrela/onwithresponse Insubsequentstepsaddpredictorswithhighestpar/alcorrela/onwithresponsecontrollingforvariablesalreadyinequa/ons

    Stopwhennumericalcriterionsignalsmaximum(minimum) Some/meseliminatevariableswhentvaluegetstoosmall

    Onlypossiblemethodforverylargepredictorpools Localop/miza/onateachstep,noguaranteeoffindingoverallop/mum

    Backwardelimina/on Startwithallpredictorsinequa/on

    Removepredictorwithsmallesttvalue Con/nueun/lnumericalcriterionsignalsmaximum(minimum)

    Oenproducesdifferentfinalmodelthanforwardstepwisemethod

  • 7/29/2019 S7_extraFeatureSelection

    6/7

    The degree of correlation between Xs.

    A high degree of multicolinearity produces unacceptable

    uncertainty (large variance) in regression coefficient estimates

    (i.e., large sampling variation)

    Imprecise estimates of slopes and even the signs of the

    coefficients may be misleading.

    t-tests which fail to reveal significant factors. Theanalysisofvariance

    fortheoverallmodelmayshowahighlysignificantlygoodfit,whenparadoxically;thetestsforindividualpredictorsarenon-significant.

    Mul/colinearity(regression)

  • 7/29/2019 S7_extraFeatureSelection

    7/7

    S7Extra:FeatureSelec/on

    ShawndraHill

    Spring2013

    TR1:30-3pmand3-4:30