feature engineering

《feature engineering》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、DiscoverFeatureEngineeringHowtoEngineerFeaturesandHowtoGetGoodatItImportanceofFeatureEngineering●Betterfeaturesmeansflexibility.●Betterfeaturesmeanssimplermodels.●Betterfeaturesmeansbetterresults.WhatisFeatureEngineering?●Featureengineeringis●theprocessoftransformingrawdataintofeatures●thatbette

2、rrepresenttheunderlyingproblemtothepredictivemodels●resultinginimprovedmodelaccuracyonunseendata.Sub-ProblemsofFeatureEngineering●FeatureImportance(correlation,randomforest)–Anestimateoftheusefulnessofafeature●FeatureExtraction(PCA)–Theautomaticconstructionofnewfeaturesfromrawdata●FeatureSelecti

3、on(rankingscore,wrapper,LASSO)–Frommanyfeaturestoafewthatareuseful●FeatureConstruction()–Themanualconstructionofnewfeaturesfromrawdata●FeatureLearning–TheautomaticidentificationanduseoffeaturesinrawdataIterativeProcessofFeatureEngineering●Brainstormfeatures●Devisefeatures●Selectfeatures●Evaluate

4、modelsGeneralExamplesofFeatureEngineering●DecomposeCategoricalAttributes–“Item_Color”thatcanbeRed,BlueorUnknown.●DecomposeaDate-Time–2014-09-20T20:45:40Z●ReframeNumericalQuantities–Num_Customer_PurchasesPurchases_Summer,Purchases_FallFeatureselectioninsklearn●Removingfeatureswithlowvariance–Vari

5、anceThreshold●Univariatefeatureselection–Regressionp-values–ClassificationAnovaF-valueVariableRanking●CorrelationCriteria–Pearsoncorrelationcoefficient●SingleVariableClassifiers–ROC(x-FPRy-TPR)AUC●InformationTheoreticRankingCriteria●Noisy(noninformative)features●Applyingunivariatefeatureselectio

6、nbeforetheSVMincreasestheSVMweightattributedtothesignificantfeaturesLimitationsofvariableranking●CanPresumablyRedundantVariablesHelpEachOther?Limitationsofvariableranking●HowDoesCorrelationImpactVariableRedundancy●Limitationsofvariableranking●CanaVariablethatisUselessbyItselfbeUsefulwithOthers?●

7、Featureselectioninsklearn●Recursivefeatureelimination–Allfeature→absoluteweightsarethesmallestarepruned(SVC)●L1-basedfeatureselection–Lasso(higheralphathefewerfeatures)–SVMsandlogistic-regression(smallerCthefewerfeatures)––●