Learning Apache Mahout Classification - DropPDF1.droppdf.com/files/0ZaSP/learning-apache-mahout-classification-by... · Learning Apache Mahout Classification Credits ... , with PDF

LearningApacheMahoutClassification

TableofContents


Credits

AbouttheAuthor

AbouttheReviewers

www.PacktPub.com

Supportfiles,eBooks,discountoffers,andmore

Whysubscribe?

FreeaccessforPacktaccountholders

Preface

Whatthisbookcovers

Whatyouneedforthisbook

Whothisbookisfor

Conventions

Readerfeedback

Customersupport

Downloadingtheexamplecode

Downloadingthecolorimagesofthisbook

Errata

Piracy

Questions

1.ClassificationinDataAnalysis

Introducingtheclassification

Applicationoftheclassificationsystem

Workingoftheclassificationsystem

Classificationalgorithms

Modelevaluationtechniques

Theconfusionmatrix

TheReceiverOperatingCharacteristics(ROC)graph

AreaundertheROCcurve

Theentropymatrix

Summary

2.ApacheMahout

IntroducingApacheMahout

AlgorithmssupportedinMahout

ReasonsforMahoutbeingagoodchoiceforclassification

InstallingMahout

BuildingMahoutfromsourceusingMaven

InstallingMaven

BuildingMahoutcode

SettingupadevelopmentenvironmentusingEclipse

SettingupMahoutforaWindowsuser

Summary

3.LearningLogisticRegression/SGDUsingMahout

Introducingregression

Understandinglinearregression

Costfunction

Gradientdescent

Logisticregression

StochasticGradientDescent

UsingMahoutforlogisticregression

Summary

4.LearningtheNaïveBayesClassificationUsingMahout

IntroducingconditionalprobabilityandtheBayesrule

UnderstandingtheNaïveBayesalgorithm

Understandingthetermsusedintextclassification

UsingtheNaïveBayesalgorithminApacheMahout

Summary

5.LearningtheHiddenMarkovModelUsingMahout

Deterministicandnondeterministicpatterns

TheMarkovprocess

IntroducingtheHiddenMarkovModel

UsingMahoutfortheHiddenMarkovModel

Summary

6.LearningRandomForestUsingMahout

Decisiontree

Randomforest

UsingMahoutforRandomforest

StepstousetheRandomforestalgorithminMahout

Summary

7.LearningMultilayerPerceptronUsingMahout

Neuralnetworkandneurons

MultilayerPerceptron

MLPimplementationinMahout

UsingMahoutforMLP

StepstousetheMLPalgorithminMahout

Summary

8.MahoutChangesintheUpcomingRelease

Mahoutnewchanges

MahoutScalaandSparkbindings

ApacheSpark

UsingMahout’sSparkshell

H2Oplatformintegration

Summary

9.BuildinganE-mailClassificationSystemUsingApacheMahout

Spame-maildataset

CreatingthemodelusingtheAssassindataset

Programtouseaclassifiermodel

Testingtheprogram

Secondusecaseasanexercise

TheASFe-maildataset

Classifierstuning

Summary

Index


LearningApacheMahoutClassificationCopyright©2015PacktPublishing

Allrightsreserved.Nopartofthisbookmaybereproduced,storedinaretrievalsystem,ortransmittedinanyformorbyanymeans,withoutthepriorwrittenpermissionofthepublisher,exceptinthecaseofbriefquotationsembeddedincriticalarticlesorreviews.

Everyefforthasbeenmadeinthepreparationofthisbooktoensuretheaccuracyoftheinformationpresented.However,theinformationcontainedinthisbookissoldwithoutwarranty,eitherexpressorimplied.Neithertheauthor,norPacktPublishing,anditsdealersanddistributorswillbeheldliableforanydamagescausedorallegedtobecauseddirectlyorindirectlybythisbook.

PacktPublishinghasendeavoredtoprovidetrademarkinformationaboutallofthecompaniesandproductsmentionedinthisbookbytheappropriateuseofcapitals.However,PacktPublishingcannotguaranteetheaccuracyofthisinformation.

Firstpublished:February2015

Productionreference:1210215

PublishedbyPacktPublishingLtd.

LiveryPlace

35LiveryStreet

BirminghamB32PB,UK.

ISBN978-1-78355-495-9

www.packtpub.com

http://www.packtpub.com

CreditsAuthor

AshishGupta

Reviewers

SivaPrakash

TharinduRusira

VishnuViswanath

CommissioningEditor

AkramHussain

AcquisitionEditor

ReshmaRaman

ContentDevelopmentEditor

MerwynD’souza

TechnicalEditors

MonicaJohn

NovinaKewalramani

ShrutiRawool

CopyEditors

SarangChari

GladsonMonteiro

AartiSaldanha

RashmiSawant

ProjectCoordinator

NehaBhatnagar

Proofreaders

SimranBhogal

SteveMaguire

Indexer

MonicaAjmeraMehta

Graphics

SheetalAute

AbhinashSahu

ProductionCoordinator

ConidonMiranda

CoverWork

ConidonMiranda

AbouttheAuthorAshishGuptahasbeenworkinginthefieldofsoftwaredevelopmentforthelast8years.Hehasworkedindifferentcompanies,suchasSAPLabsandCaterpillar,asasoftwaredeveloper.Whileworkingforastart-upwherehewasresponsibleforpredictingpotentialcustomersfornewfashionapparelsusingsocialmedia,hedevelopedaninterestinthefieldofmachinelearning.Sincethen,hehasworkedonusingbigdatatechnologiesandmachinelearningfordifferentindustries,includingretail,finance,insurance,andsoon.Hehasapassionforlearningnewtechnologiesandsharingtheknowledgethusgainedwithothers.HehasorganizedmanybootcampsfortheApacheMahoutandHadoopecosystem.

Firstofall,Iwouldliketothankopensourcecommunitiesfortheircontinuouseffortsindevelopinggreatsoftwareforall.IwouldliketothankMerwynD’SouzaandReshmaRaman,myeditorsforthisproject.Specialthankstothereviewersofthisbook.

Nothingcanbeaccomplishedwithoutthesupportoffamily,friends,andlovedones.Iwouldliketothankmyfriends,family,andespeciallymywifeandmysonfortheircontinuoussupportthroughoutthewritingofthisbook.

AbouttheReviewersSivaPrakashisworkingasatechleadinBangalore.Hehasextensivedevelopmentexperienceintheanalysis,design,development,implementation,andmaintenanceofvariousdesktop,mobile,andweb-basedapplications.Helovestrekking,traveling,music,readingbooks,andblogging.

YoucanfindhimonLinkedInathttps://www.linkedin.com/in/techsivam.

TharinduRusiraiscurrentlyacomputerscienceandengineeringundergraduateattheUniversityofMoratuwa,SriLanka.Asastudentresearcher,hehasstronginterestsinmachinelearning,compilers,andhigh-performancecomputing.

TharinduhasalsoworkedasaresearchanddevelopmentsoftwareengineeringinternatZaiziAsia(Pvt)Ltd.,wherehefirststartedusingApacheMahoutduringtheimplementationofanenterprise-levelcontentmanagementandinformationretrievalsystem.

HeseesthepotentialofApacheMahoutasascalablemachinelearninglibraryforindustry-levelimplementationsandhasevencontributedtotheMahout0.9release,thelateststablereleaseofMahout.

HeisavailableonLinkedInathttps://www.linkedin.com/in/trusira.

VishnuViswanathisaseniorbigdatadeveloperwhohasmanyyearsofindustrialexpertiseinthearenaofmachinelearning.Heisatechenthusiastandispassionateaboutbigdataandhasexpertiseonmostbig-data-relatedtechnologies.

YoucanfindhimonLinkedInathttp://in.linkedin.com/in/vishnuviswanath25.

https://www.linkedin.com/in/techsivam

https://www.linkedin.com/in/trusira

http://in.linkedin.com/in/vishnuviswanath25

www.PacktPub.com

Supportfiles,eBooks,discountoffers,andmoreForsupportfilesanddownloadsrelatedtoyourbook,pleasevisitwww.PacktPub.com.

DidyouknowthatPacktofferseBookversionsofeverybookpublished,withPDFandePubfilesavailable?YoucanupgradetotheeBookversionatwww.PacktPub.comandasaprintbookcustomer,youareentitledtoadiscountontheeBookcopy.Getintouchwithusat<[email protected]>formoredetails.

Atwww.PacktPub.com,youcanalsoreadacollectionoffreetechnicalarticles,signupforarangeoffreenewslettersandreceiveexclusivediscountsandoffersonPacktbooksandeBooks.

https://www2.packtpub.com/books/subscription/packtlib

DoyouneedinstantsolutionstoyourITquestions?PacktLibisPackt’sonlinedigitalbooklibrary.Here,youcansearch,access,andreadPackt’sentirelibraryofbooks.

http://www.PacktPub.com


mailto:[email protected]


https://www2.packtpub.com/books/subscription/packtlib

Whysubscribe?FullysearchableacrosseverybookpublishedbyPacktCopyandpaste,print,andbookmarkcontentOndemandandaccessibleviaawebbrowser

FreeaccessforPacktaccountholdersIfyouhaveanaccountwithPacktatwww.PacktPub.com,youcanusethistoaccessPacktLibtodayandview9entirelyfreebooks.Simplyuseyourlogincredentialsforimmediateaccess.


PrefaceThankstotheprogressmadeinthehardwareindustries,ourstoragecapacityhasincreased,andbecauseofthis,therearemanyorganizationswhowanttostorealltypesofeventsforanalyticspurposes.Thishasgivenbirthtoaneweraofmachinelearning.Thefieldofmachinelearningisverycomplexandwritingthesealgorithmsisnotapieceofcake.ApacheMahoutprovidesuswithreadymadealgorithmsintheareaofmachinelearningandsavesusfromthecomplextaskofalgorithmimplementation.

TheintentionofthisbookistocoverclassificationalgorithmsavailableinApacheMahout.Whetheryouhavealreadyworkedonclassificationalgorithmsusingsomeothertoolorarecompletelynewtothefield,thisbookwillhelpyou.So,startreadingthisbooktoexploretheclassificationalgorithmsinoneofthemostpopularopensourceprojectswhichenjoysstrongcommunitysupport:ApacheMahout.

WhatthisbookcoversChapter1,ClassificationinDataAnalysis,providesanintroductiontotheclassificationconceptindataanalysis.Thischapterwillcoverthebasicsofclassification,similaritymatrix,andalgorithmsavailableinthisarea.

Chapter2,ApacheMahout,providesanintroductiontoApacheMahoutanditsinstallationprocess.Further,thischapterwilltalkaboutwhyitisagoodchoiceforclassification.

Chapter3,LearningLogisticRegression/SGDUsingMahout,discusseslogisticregressionandStochasticGradientDescent,andhowdeveloperscanuseMahouttouseSGD.

Chapter4,LearningtheNaïveBayesClassificationUsingMahout,discussestheBayesTheorem,NaïveBayesclassification,andhowwecanuseMahouttobuildNaïveBayesclassifier.

Chapter5,LearningtheHiddenMarkovModelUsingMahout,coverstheHMMandhowtouseMahout’sHMMalgorithms.

Chapter6,LearningRandomForestUsingMahout,discussestheRandomforestalgorithmindetail,andhowtouseMahout’sRandomforestimplementation.

Chapter7,LearningMultilayerPerceptronUsingMahout,discussesMahoutasanearlylevelimplementationofaneuralnetwork.WewilldiscussMultilayerPerceptroninthischapter.Further,wewilluseMahout’simplementationofMLP.

Chapter8,MahoutChangesintheUpcomingRelease,discussesMahoutasaworkinprogress.WewilldiscussthenewmajorchangesintheupcomingreleaseofMahout.

Chapter9,BuildinganE-mailClassificationSystemUsingApacheMahout,providestwousecasesofe-mailclassification—spammailclassificationande-mailclassificationbasedontheprojectthemailbelongsto.Wewillcreatethemodel,andusethismodelinaprogramthatwillsimulatetherealworkingenvironment.

WhatyouneedforthisbookTousetheexamplesinthisbook,youshouldhavethefollowingsoftwareinstalledonyoursystem:

Java1.6orhigherEclipseHadoopMahout;wewilldiscusstheinstallationinChapter2,ApacheMahout,ofthisbookMaven,dependingonhowyouinstallMahout

WhothisbookisforIfyouareadatascientistwhohassomeexperiencewiththeHadoopecosystemandmachinelearningmethodsandwanttotryoutclassificationonlargedatasetsusingMahout,thisbookisidealforyou.KnowledgeofJavaisessential.

ConventionsInthisbook,youwillfindanumberoftextstylesthatdistinguishbetweendifferentkindsofinformation.Herearesomeexamplesofthesestylesandanexplanationoftheirmeaning.

Codewordsintext,databasetablenames,foldernames,filenames,fileextensions,pathnames,dummyURLs,userinput,andTwitterhandlesareshownasfollows:“Extractthesourcecodeandensurethatthefoldercontainsthepom.xmlfile.”

Ablockofcodeissetasfollows:

publicstaticMap<String,Integer>readDictionary(Configurationconf,

PathdictionaryPath){

Map<String,Integer>dictionary=newHashMap<String,Integer>();

for(Pair<Text,IntWritable>pair:newSequenceFileIterable<Text,

IntWritable>(dictionaryPath,true,conf)){

dictionary.put(pair.getFirst().toString(),

pair.getSecond().get());

}

returndictionary;

}

Whenwewishtodrawyourattentiontoaparticularpartofacodeblock,therelevantlinesoritemsaresetinbold:






dictionary.put(pair.getFirst().toString(),


}

returndictionary;

}

Anycommand-lineinputoroutputiswrittenasfollows:

hadoopfs-mkdir/user/hue/KDDTrain

hadoopfs-mkdir/user/hue/KDDTest

hadoopfs–put/tmp/KDDTrain+_20Percent.arff/user/hue/KDDTrain

hadoopfs–put/tmp/KDDTest+.arff/user/hue/KDDTest

Newtermsandimportantwordsareshowninbold.Wordsthatyouseeonthescreen,forexample,inmenusordialogboxes,appearinthetextlikethis:“Now,navigatetothelocationformahout-distribution-0.9andclickonFinish.”

NoteWarningsorimportantnotesappearinaboxlikethis.

TipTipsandtricksappearlikethis.

ReaderfeedbackFeedbackfromourreadersisalwayswelcome.Letusknowwhatyouthinkaboutthisbook—whatyoulikedordisliked.Readerfeedbackisimportantforusasithelpsusdeveloptitlesthatyouwillreallygetthemostoutof.

Tosendusgeneralfeedback,simplye-mail<[email protected]>,andmentionthebook’stitleinthesubjectofyourmessage.

Ifthereisatopicthatyouhaveexpertiseinandyouareinterestedineitherwritingorcontributingtoabook,seeourauthorguideatwww.packtpub.com/authors.


http://www.packtpub.com/authors

CustomersupportNowthatyouaretheproudownerofaPacktbook,wehaveanumberofthingstohelpyoutogetthemostfromyourpurchase.

DownloadingtheexamplecodeYoucandownloadtheexamplecodefilesfromyouraccountathttp://www.packtpub.comforallthePacktPublishingbooksyouhavepurchased.Ifyoupurchasedthisbookelsewhere,youcanvisithttp://www.packtpub.com/supportandregistertohavethefilese-maileddirectlytoyou.


http://www.packtpub.com/support

DownloadingthecolorimagesofthisbookWealsoprovideyouwithaPDFfilethathascolorimagesofthescreenshots/diagramsusedinthisbook.Thecolorimageswillhelpyoubetterunderstandthechangesintheoutput.Youcandownloadthisfilefromhttp://www.packtpub.com/sites/default/files/downloads/4959OS_ColoredImages.pdf.

http://www.packtpub.com/sites/default/files/downloads/4959OS_ColoredImages.pdf

ErrataAlthoughwehavetakeneverycaretoensuretheaccuracyofourcontent,mistakesdohappen.Ifyoufindamistakeinoneofourbooks—maybeamistakeinthetextorthecode—wewouldbegratefulifyoucouldreportthistous.Bydoingso,youcansaveotherreadersfromfrustrationandhelpusimprovesubsequentversionsofthisbook.Ifyoufindanyerrata,pleasereportthembyvisitinghttp://www.packtpub.com/submit-errata,selectingyourbook,clickingontheErrataSubmissionFormlink,andenteringthedetailsofyourerrata.Onceyourerrataareverified,yoursubmissionwillbeacceptedandtheerratawillbeuploadedtoourwebsiteoraddedtoanylistofexistingerrataundertheErratasectionofthattitle.

Toviewthepreviouslysubmittederrata,gotohttps://www.packtpub.com/books/content/supportandenterthenameofthebookinthesearchfield.TherequiredinformationwillappearundertheErratasection.

http://www.packtpub.com/submit-errata

https://www.packtpub.com/books/content/support

PiracyPiracyofcopyrightedmaterialontheInternetisanongoingproblemacrossallmedia.AtPackt,wetaketheprotectionofourcopyrightandlicensesveryseriously.IfyoucomeacrossanyillegalcopiesofourworksinanyformontheInternet,pleaseprovideuswiththelocationaddressorwebsitenameimmediatelysothatwecanpursuearemedy.

Pleasecontactusat<[email protected]>withalinktothesuspectedpiratedmaterial.

Weappreciateyourhelpinprotectingourauthorsandourabilitytobringyouvaluablecontent.


QuestionsIfyouhaveaproblemwithanyaspectofthisbook,youcancontactusat<[email protected]>,andwewilldoourbesttoaddresstheproblem.


Chapter1.ClassificationinDataAnalysisInthelastdecade,wesawahugegrowthinsocialnetworkingande-commercesites.IamsurethatyoumusthavegotinformationaboutthisbookonFacebook,Twitter,orsomeothersite.Chancesarealsohighthatyouarereadingane-copyofthisbookafterorderingitonyourphoneortablet.

ThismustgiveyouanideaofhowmuchdatawearegeneratingovertheInterneteverysingleday.Now,inordertoobtainallnecessaryinformationfromthedata,wenotonlycreatedatabutalsostorethisdata.Thisdataisextremelyusefultogetsomeimportantinsightsintothebusiness.Theanalysisofthisdatacanincreasethecustomerbaseandcreateprofitsfortheorganization.Taketheexampleofane-commercesite.Youvisitthesitetobuysomebook.Yougetinformationaboutbooksonrelatedtopicsorthesametopic,publisher,orwriter,andthishelpsyoutotakebetterdecisions,whichalsohelpsthesitetoknowmoreaboutitscustomers.Thiswilleventuallyleadtoanincreaseinsales.

Findingrelateditemsorsuggestinganewitemtotheuserisallpartofthedatascienceinwhichweanalyzethedataandtrytogetusefulpatterns.

Dataanalysisistheprocessofinspectinghistoricaldataandcreatingmodelstogetusefulinformationthatisrequiredtohelpindecisionmaking.Itishelpfulinmanyindustries,suchase-commerce,banking,finance,healthcare,telecommunications,retail,oceanography,andmanymore.

Let’staketheexampleofaweatherforecastingsystem.Itisasystemthatcanpredictthestateoftheatmosphereataparticularlocation.Inthisprocess,scientistscollecthistoricaldataoftheatmosphereofthatlocationandtrytocreateamodelbasedonittopredicthowtheatmospherewillevolveoveraperiodoftime.

Inmachinelearning,classificationistheautomationofthedecision-makingprocessthatlearnsfromexamplesofthepastandemulatesthosedecisionsautomatically.Emulatingthedecisionsautomaticallyisacoreconceptinpredictiveanalytics.Inthischapter,wewilllookatthefollowingpoints:

UnderstandingclassificationWorkingofclassificationsystemsClassificationalgorithmsModelevaluationmethods

IntroducingtheclassificationThewordclassificationalwaysremindsusofourbiologyclass,wherewelearnedabouttheclassificationofanimals.Welearnedaboutdifferentcategoriesofanimals,suchasmammals,reptiles,birds,amphibians,andsoon.

Ifyourememberhowthesecategoriesaredefined,youwillrealizethattherewerecertainpropertiesthatscientistsfoundinexistinganimals,andbasedontheseproperties,theycategorizedanewanimal.

Otherreal-lifeexamplesofclassificationcouldbe,forinstance,whenyouvisitthedoctor.He/sheasksyoucertainquestions,andbasedonyouranswers,he/sheisabletoidentifywhetheryouhaveacertaindiseaseornot.

Classificationisthecategorizationofpotentialanswers,andinmachinelearning,wewanttoautomatethisprocess.Biologicalclassificationisanexampleofmulticlassclassificationandfindingthediseaseisanexampleofbinaryclassification.

Indataanalysis,wewanttousemachinelearningconcepts.Toanalyzethedata,wewanttobuildasystemthatcanhelpustofindoutwhichclassanindividualitembelongsto.Usually,theseclassesaremutuallyexclusive.Arelatedprobleminthisareaisfindingouttheprobabilitythatanindividualbelongstoacertainclass.

Classificationisasupervisedlearningtechnique.Inthistechnique,machines—basedonhistoricaldata—learnandgainthecapabilitiestopredicttheunknown.Inmachinelearning,anotherpopulartechniqueisunsupervisedlearning.Insupervisedlearning,wealreadyknowtheoutputcategories,butinunsupervisedlearning,weknownothingabouttheoutput.Let’sunderstandthiswithaquickexample:supposewehaveafruitbasket,andwewanttoclassifyfruits.Whenwesayclassify,itmeansthatinthetrainingdata,wealreadyhaveoutputvariables,suchassizeandcolor,andweknowwhetherthecolorisredandthesizeisfrom2.3”to3.7”.Wewillclassifythatfruitasanapple.Oppositetothis,inunsupervisedlearning,wewanttoseparatedifferentfruits,andwedonothaveanyoutputinformationinthetrainingdataset,sothelearningalgorithmwillseparatedifferentfruitsbasedondifferentfeaturespresentinthedataset,butitwillnotbeabletolabelthem.Inotherwords,itwillnotbeabletotellwhichoneisanappleandwhichoneisabanana,althoughitwillbeabletoseparatethem.

ApplicationoftheclassificationsystemClassificationisusedforprediction.Inthecaseofe-mailcategorization,itisusedtoclassifye-mailasspamornotspam.Nowadays,Gmailisclassifyinge-mailsasprimary,social,andpromotionalaswell.Classificationisusefulinpredictingcreditcardfrauds,tocategorizecustomersforeligibilityofloans,andsoon.Itisalsousedtopredictcustomerchurnintheinsuranceandtelecomindustries.Itisusefulinthehealthcareindustryaswell.Basedonhistoricaldata,itisusefulinclassifyingparticularsymptomsofadiseasetopredictthediseaseinadvance.Classificationcanbeusedtoclassifytropicalcyclones.So,itisusefulacrossallindustries.

WorkingoftheclassificationsystemLet’sunderstandtheclassificationprocessinmoredetail.Intheprocessofclassification,withthedatasetgiventous,wetrytofindoutinformativevariablesusingwhichwecanreducetheuncertaintyandcategorizesomething.Theseinformativevariablesarecalledexplanatoryvariablesorfeatures.

Thefinalcategoriesthatweareinterestedarecalledtargetvariablesorlabels.Explanatoryvariablescanbeanyofthefollowingforms:

Continuous(numerictypes)CategoricalWord-likeText-like

NoteIfnumerictypesarenotusefulforanymathematicalfunctions,thosewillbecountedascategorical(zipcodes,streetnumbers,andsoon).

So,forexample,wehaveadatasetofcustomer’s’loanapplications,andwewanttobuildaclassifiertofindoutwhetheranewcustomeriseligibleforaloanornot.Inthisdataset,wecanhavethefollowingfields:

CustomerAgeCustomerIncome(PA)CustomerAccountBalanceLoanGranted

Fromthesefields,CustomerAge,CustomerIncome(PA)andCustomerAccountBalancewillworkasexplanatoryvariablesandLoanGrantedwillbethetargetvariable,asshowninthefollowingscreenshot:

Tounderstandthecreationoftheclassifier,weneedtounderstandafewterms,asshowninthefollowingdiagram:

Trainingdataset:Fromthegivendataset,aportionofthedataisusedtocreatethetrainingdataset(itcouldbe70percentofthegivendata).Thisdatasetisusedtobuildtheclassifier.Allthefeaturesetsareusedinthisdataset.Testdataset:Thedatasetthatisleftafterthetrainingdatasetisusedtotestthecreatedmodel.Withthisdata,onlythefeaturesetisusedandthemodelisusedtopredictthetargetvariablesorlabels.Model:Thisisusedtounderstandthealgorithmusedtogeneratethetargetvariables.

Whilebuildingaclassifier,wefollowthesesteps:

CollectinghistoricaldataCleaningdata(alotofactivitiesareinvolvedhere,suchasspaceremoval,andsoon)DefiningtargetvariablesDefiningexplanatoryvariablesSelectinganalgorithmTrainingthemodel(usingthetrainingdataset)RunningtestdataEvaluatingthemodelAdjustingexplanatoryvariablesRerunningthetest

Whilepreparingthemodel,oneshouldtakecareofoutlierdetection.Outlierdetectionisamethodtofindoutitemsthatdonotconformtoanexpectedpatterninadataset.Outliersinaninputdatasetcanmisleadthetrainingprocessofanalgorithm.Thiscanaffectthemodelaccuracy.Therearealgorithmstofindouttheseoutliersinthedatasets.Distance-

basedtechniquesandfuzzy-logic-basedmethodsaremostlyusedtofindoutoutliersinthedataset.Let’stalkaboutoneexampletounderstandtheoutliers.

Wehaveasetofnumbers,andwewanttofindoutthemeanofthesenumbers:

10,75,10,15,20,85,25,30,25

Justplotthesenumbersandtheresultwillbeasshowninthefollowingscreenshot:

Clearly,thenumbers75and85areoutliers(farawayintheplotfromtheothernumbers).

Mean=sumofvalues/numberofvalues=32.78

Meanwithouttheoutliers:=19.29

So,nowyoucanunderstandhowoutlierscanaffecttheresults.

Whilecreatingthemodel,wecanencountertwomajorlyoccurringproblems—OverfittingandUnderfitting.

Overfittingoccurswhenthealgorithmcapturesthenoiseofthedata,andthealgorithmfitsthedatatoowell.Generally,itoccursifweuseallthegivendatatobuildthemodelusingpurememorization.Insteadoffindingoutthegeneralizingpattern,themodeljustmemorizesthepattern.Usually,inthecaseofoverfitting,themodelgetsmorecomplex,anditisallowedtopickupspuriouscorrelations.Thesecorrelationsarespecifictotrainingdatasetsanddonotrepresentcharacteristicsofthewholedatasetingeneral.

Thefollowingdiagramisanexampleofoverfitting.Anoutlierispresent,andthealgorithmconsidersthatandcreatesamodelthatperfectlyclassifiesthetrainingset,butbecauseofthis,thetestdataiswronglyclassified(boththerectanglesareclassifiedasstarsinthetestdata):

Thereisnosinglemethodtoavoidoverfitting;however,wehavesomeapproaches,suchasareductioninthenumberoffeaturesandtheregularizationofafewofthefeatures.Anotherwayistotrainthemodelwithsomedatasetandtestwiththeremainingdataset.Acommonmethodcalledcross-validationisusedtogeneratemultipleperformancemeasures.Inthisway,asingledatasetissplitandusedforthecreationofperformancemeasures.

Underfittingoccurswhenthealgorithmcannotcapturethepatternsinthedata,andthedatadoesnotfitwell.Underfittingisalsoknownashighbias.Itmeansyouralgorithmhassuchastrongbiastowardsitshypothesisthatitdoesnotfitthedatawell.Foranunderfittingerror,moredatawillnothelp.Itcanincreasethetrainingerror.Moreexplanatoryvariablescanhelptodealwiththeunderfittingproblem.Moreexplanatoryfieldswillexpandthehypothesisspaceandwillbeusefultoovercomethisproblem.

Bothoverfittingandunderfittingprovidepoorresultswithnewdatasets.

ClassificationalgorithmsWewillnowdiscussthefollowingalgorithmsthataresupportedbyApacheMahoutinthisbook:

Logisticregression/StochasticGradientDescent(SGD):Weusuallyreadregressionalongwithclassification,butactually,thereisadifferencebetweenthetwo.Classificationinvolvesacategoricaltargetvariable,whileregressioninvolvesanumerictargetvariable.Classificationpredictswhethersomethingwillhappen,andregressionpredictshowmuchofsomethingwillhappen.WewillcoverthisalgorithminChapter3,LearningLogisticRegression/SGDUsingMahout.MahoutsupportslogisticregressiontrainedviaStochasticGradientDescent.NaïveBayesclassification:Thisisaverypopularalgorithmfortextclassification.NaïveBayesusestheconceptofprobabilitytoclassifynewitems.ItisbasedontheBayestheorem.WewilldiscussthisalgorithminChapter4,LearningtheNaïveBayesClassificationUsingMahout.Inthischapter,wewillseehowMahoutisusefulinclassifyingtext,whichisrequiredinthedataanalysisfield.Wewilldiscussvectorization,bagofwords,n-grams,andothertermsusedintextclassification.HiddenMarkovModel(HMM):Thisisusedinvariousfields,suchasspeechrecognition,parts-of-speechtagging,geneprediction,time-seriesanalysis,andsoon.InHMM,weobserveasequenceofemissionsbutdonothaveasequenceofstateswhichamodelusestogeneratetheemission.InChapter5,LearningtheHiddenMarkovModelUsingMahout,wewilltakeonemorealgorithmsupportedbyMahoutHiddenMarkovModel.WewilldiscussHMMindetailandseehowMahoutsupportsthisalgorithm.RandomForest:Thisisthemostwidelyusedalgorithminclassification.RandomForestconsistsofacollectionofsimpletreepredictors,eachcapableofproducingaresponsewhenpresentedwithasetofexplanatoryvariables.InChapter6,LearningRandomForestUsingMahout,wewilldiscussthisalgorithmindetailandalsotalkabouthowtouseMahouttoimplementthisalgorithm.Multi-layerPerceptron(MLP):InChapter7,LearningMultilayerPerceptronUsingMahout,wewilldiscussthisnewlyimplementedalgorithminMahout.AnMLPconsistsofmultiplelayersofnodesinadirectedgraph,witheachlayerfullyconnectedtothenextone.Itisabasefortheimplementationofneuralnetworks.WewilldiscussneuralnetworksalittlebutonlyafteradetaileddiscussiononMLPinMahout.

WewilldiscussalltheclassificationalgorithmssupportedbyApacheMahoutinthisbook,andwewillalsocheckthemodelevaluationtechniquesprovidedbyApacheMahout.

ModelevaluationtechniquesWecannothaveasingleevaluationmetricthatcanfitalltheclassifiermodels,butwecanfindoutsomecommonissuesinevaluation,andwehavetechniquestodealwiththem.WewilldiscussthefollowingtechniquesthatareusedinMahout:

ConfusionmatrixROCgraphAUCEntropymatrix

TheconfusionmatrixTheconfusionmatrixprovidesuswiththenumberofcorrectandincorrectpredictionsmadebythemodelcomparedwiththeactualoutcomes(targetvalues)inthedata.AconfusionmatrixisaN*Nmatrix,whereNisthenumberoflabels(classes).Eachcolumnisaninstanceinthepredictedclass,andeachrowisaninstanceintheactualclass.Usingthismatrix,wecanfindouthowoneclassisconfusedwithanother.Let’sassumethatwehaveaclassifierthatclassifiesthreefruits:strawberries,cherries,andgrapes.Assumingthatwehaveasampleof24fruits:7strawberries,8cherries,and9grapes,theresultingconfusionmatrixwillbeasshowninthefollowingtable:

Predictedclassesbymodel

Actualclass

Strawberries Cherries Grapes

Strawberries 4 3 0

Cherries 2 5 1

Grapes 0 1 8

So,inthismodel,fromthe8strawberries,3wereclassifiedascherries.Fromthe8cherries,2wereclassifiedasstrawberries,and1isclassifiedasagrape.Fromthe9grapes,1isclassifiedasacherry.Fromthismatrix,wewillcreatethetableofconfusion.Thetableofconfusionhastworowsandtwocolumnsthatreportabouttruepositive,truenegative,falsepositive,andfalsenegative.

So,ifwebuildthistableforaparticularclass,let’ssayforstrawberries,itwouldbeasfollows:

TruePositive

4(actualstrawberriesclassifiedcorrectly)(a)

FalsePositive

2(cherriesthatwereclassifiedasstrawberries)(b)

FalseNegative

3(strawberrieswronglyclassifiedascherries)(c)

TrueNegative

15(allotherfruitscorrectlynotclassifiedasstrawberries)(d)

Usingthistableofconfusion,wecanfindoutthefollowingterms:

Accuracy:Thisistheproportionofthetotalnumberofpredictionsthatwerecorrectlyclassified.Itiscalculatedas(TruePositive+TrueNegative)/Positive+Negative.Therefore,accuracy=(a+d)/(a+b+c+d).Precisionorpositivepredictivevalue:Thisistheproportionofpositivecasesthatwerecorrectlyclassified.Itiscalculatedas(TruePositive)/(TruePositive+FalsePositive).Therefore,precision=a/(a+b).Negativepredictivevalue:Thisistheproportionofnegativecasesthatwereclassifiedcorrectly.ItiscalculatedasTrueNegative/(TrueNegative+FalseNegative).Therefore,negativepredictivevalue=d/(c+d).Sensitivity/truepositiverate/recall:Thisistheproportionoftheactualpositive

casesthatwerecorrectlyidentified.ItiscalculatedasTruePositive/(TruePositive+FalseNegative).Therefore,sensitivity=a/(a+c).Specificity:Thisistheproportionoftheactualnegativecases.ItiscalculatedasTrueNegative/(FalsePositive+TrueNegative).Therefore,specificity=d/(b+d).F1score:Thisisthemeasureofatest’saccuracy,anditiscalculatedasfollows:F1=2.((Positivepredictivevalue(precision)*sensitivity(recall))/(Positivepredictivevalue(precision)+sensitivity(recall))).

TheReceiverOperatingCharacteristics(ROC)graphROCisatwo-dimensionalplotofaclassifierwithfalsepositiverateonthexaxisandtruepositiverateontheyaxis.Thelowerpoint(0,0)inthefigurerepresentsneverissuingapositiveclassification.Point(0,1)representsperfectclassification.Thediagonalfrom(0,0)to(1,1)dividestheROCspace.Pointsabovethediagonalrepresentgoodclassificationresults,andpointsbelowthelinerepresentpoorresults,asshowninthefollowingdiagram:

AreaundertheROCcurveThisistheareaundertheROCcurveandisalsoknownasAUC.Itisusedtomeasurethequalityoftheclassificationmodel.Inpractice,mostoftheclassificationmodelshaveanAUCbetween0.5and1.Thecloserthevalueisto1,thegreaterisyourclassifier.

TheentropymatrixBeforegoingintothedetailsoftheentropymatrix,firstweneedtounderstandentropy.TheconceptofentropyininformationtheorywasdevelopedbyShannon.

Entropyisameasureofdisorderthatcanbeappliedtoaset.Itisdefinedas:

Entropy=-p1log(p1)–p2log(p2)-…….

Eachpistheprobabilityofaparticularpropertywithintheset.Let’srevisitourcustomerloanapplicationdataset.Forexample,assumingwehaveasetof10customersfromwhich6areeligibleforaloanand4arenot.Here,wehavetwoproperties(classes):eligibleornoteligible.

P(eligible)=6/10=0.6

P(noteligible)=4/10=0.4

So,entropyofthedatasetwillbe:

Entropy=-[0.6*log2(0.6)+0.4*log2(0.4)]

=-[0.6*-0.74+0.4*-1.32]

=0.972

Entropyisusefulinacquiringknowledgeofinformationgain.Informationgainmeasuresthechangeinentropyduetoanynewinformationbeingaddedinmodelcreation.So,ifentropydecreasesfromnewinformation,itindicatesthatthemodelisperformingwellnow.Informationgainiscalculatedas:

IG(classes,subclasses)=entropy(class)–(p(subclass1)*entropy(subclass1)+p(subclass2)*entropy(subclass2)+…)

Entropymatrixisbasicallythesameastheconfusionmatrixdefinedearlier;theonlydifferenceisthattheelementsinthematrixaretheaveragesofthelogoftheprobabilityscoreforeachtrueorestimatedcategorycombination.Agoodmodelwillhavesmallnegativenumbersalongthediagonalandwillhavelargenegativenumbersintheoff-diagonalposition.

SummaryWehavediscussedclassificationanditsapplicationsandalsowhatalgorithmandclassifierevaluationtechniquesaresupportedbyMahout.Wediscussedtechniqueslikeconfusionmatrix,ROCgraph,AUC,andentropymatrix.

Now,wewillmovetothenextchapterandsetupApacheMahoutandthedeveloperenvironment.WewillalsodiscussthearchitectureofApacheMahoutandfindoutwhyMahoutisagoodchoiceforclassification.

Chapter2.ApacheMahoutInthepreviouschapter,wediscussedclassificationandlookedintothealgorithmsprovidedbyMahoutinthisarea.Beforegoingtothosealgorithms,weneedtounderstandMahoutanditsinstallation.Inthischapter,wewillexplorethefollowingtopics:

WhatisApacheMahout?AlgorithmssupportedinMahoutWhyisitagoodchoiceforclassificationproblems?SettingupthesystemforMahoutdevelopment

IntroducingApacheMahoutAmahoutisapersonwhoridesandcontrolsanelephant.MostofthealgorithmsinApacheMahoutareimplementedontopofHadoop,whichisanotherApache-licensedprojectandhasthesymbolofanelephant(http://hadoop.apache.org/).AsApacheMahoutridesoverHadoop,thisnameisjustified.

ApacheMahoutisaprojectofApacheSoftwareFoundationthathasimplementationsofmachinelearningalgorithms.MahoutwasstartedasasubprojectoftheApacheLuceneprojectin2008.Aftersometime,anopensourceprojectnamedTaste,whichwasdevelopedforcollaborativefiltering,anditwasabsorbedintoMahout.MahoutiswritteninJavaandprovidesscalablemachinelearningalgorithms.Mahoutisthedefaultchoiceformachinelearningproblemsinwhichthedataistoolargetofitintoasinglemachine.MahoutprovidesJavalibrariesanddoesnotprovideanyuserinterfaceorserver.Itisaframeworkoftoolstobeusedandadaptedbydevelopers.

Tosumitup,Mahoutprovidesyouwithimplementationsofthemostfrequentlyusedmachinelearningalgorithmsintheareaofclassification,clustering,andrecommendation.Insteadofspendingtimewritingalgorithms,itprovidesuswithready-to-consumesolutions.

MahoutusesHadoopforitsalgorithms,butsomeofthealgorithmscanalsorunwithoutHadoop.Currently,Mahoutsupportsthefollowingusecases:

Recommendation:Thistakestheuserdataandtriestopredictitemsthattheusermightlike.Withthisusecase,youcanseeallthesitesthataresellinggoodstotheuser.Basedonyourpreviousaction,theywilltrytofindoutunknownitemsthatcouldbeofuse.Oneexamplecanbethis:assoonasyouselectsomebookfromAmazon,thewebsitewillshowyoualistofotherbooksunderthetitle,CustomersWhoBoughtThisItemAlsoBought.Italsoshowsthetitle,WhatOtherItemsDoCustomersBuyAfterViewingThisItem?AnotherexampleofrecommendationisthatwhileplayingvideosonYouTube,itrecommendsthatyoulistentosomeothervideosbasedonyourselection.MahoutprovidesfullAPIsupporttodevelopyour

http://hadoop.apache.org/

ownuser-basedoritem-basedrecommendationengine.Classification:Asdefinedintheearlierchapter,classificationdecideshowmuchanitembelongstooneparticularcategory.E-mailclassificationforfilteringoutspamisaclassicexampleofclassification.MahoutprovidesarichsetofAPIstobuildyourownclassificationmodel.Forexample,Mahoutcanbeusedtobuildadocumentclassifierorane-mailclassifier.Clustering:Thisisatechniquethattriestogroupitemstogetherbasedonsomesortofsimilarity.Here,wefindthedifferentclustersofitemsbasedoncertainproperties,andwedonotknowthenameoftheclusterinadvance.Themaindifferencebetweenclusteringandclassificationisthatinclassification,weknowtheendclassname.Clusteringisusefulinfindingoutdifferentcustomersegments.GoogleNewsusestheclusteringtechniqueinordertogroupnews.Forclustering,Mahouthasalreadyimplementedsomeofthemostpopularalgorithmsinthisarea,suchask-means,fuzzyk-means,canopy,andsoon.Dimensionalreduction:Aswediscussedinthepreviouschapter,featuresarecalleddimensions.Dimensionalreductionistheprocessofreducingthenumberofrandomvariablesunderconsideration.Thismakesdataeasytouse.Mahoutprovidesalgorithmsfordimensionalreduction.SingularvaluedecompositionandLanczosareexamplesofthealgorithmsthatMahoutprovides.Topicmodeling:Topicmodelingisusedtocapturetheabstractideaofadocument.Atopicmodelisamodelthatassociatesprobabilitydistributionwitheachdocumentovertopics.Giventhatadocumentisaboutaparticulartopic,onewouldexpectparticularwordstoappearinthedocumentmoreorlessfrequently.“Football”and“goal”willappearmoreinadocumentaboutsports.LatentDirichletAllocation(LDA)isapowerfullearningalgorithmfortopicmodeling.InMahout,collapsedvariationalBayesisimplementedforLDA.

AlgorithmssupportedinMahoutTheimplementationofalgorithmsinMahoutcanbecategorizedintotwogroups:

Sequentialalgorithms:ThesealgorithmsareexecutedsequentiallyanddonotuseHadoopscalableprocessing.TheyareusuallytheonesderivedfromTaste.Forexample:user-basedcollaborativefiltering,logisticregression,HiddenMarkovModel,multi-layerperceptron,singularvaluedecomposition.Parallelalgorithms:ThesealgorithmscansupportpetabytesofdatausingHadoop’smapandhencereduceparallelprocessing.Forexample,RandomForest,NaïveBayes,canopyclustering,k-meansclustering,spectralclustering,andsoon.

ReasonsforMahoutbeingagoodchoiceforclassificationInmachinelearningsystems,themoredatayouuse,themoreaccuratethesystembuiltwillbe.Mahout,whichusesHadoopforscalability,iswayaheadofothersintermsofhandlinghugedatasets.Asthenumberoftrainingsetsincreases,Mahout’sperformancealsoincreases.Iftheinputsizefortrainingexamplesisfrom1millionto10million,thenMahoutisanexcellentchoice.

Forclassificationproblems,increaseddatafortrainingisdesirableasitcanimprovetheaccuracyofthemodel.Generally,asthenumberofdatasetsincreases,memoryrequirementalsoincreases,andalgorithmsbecomeslow,butMahout’sscalableandparallelalgorithmsworkbetterwithregardstothetimetaken.Eachnewmachineaddeddecreasesthetrainingtimeandprovideshigherperformance.

InstallingMahoutNowlet’strytheslightlychallengingpartofthisbook:Mahoutinstallation.Basedoncommonexperiences,Ihavecomeupwiththefollowingquestionsorconcernsthatusersfacebeforeinstallation:

IdonotknowanythingaboutMaven.HowwillIcompileMahoutbuild?HowcanIsetupEclipsetowritemyownprogramsinMahout?HowcanIinstallMahoutonaWindowssystem?

So,wewillinstallMahoutwiththehelpofthefollowingsteps.Eachstepisindependentfromtheother.Youcanchooseanyoneofthese:

BuildingMahoutcodeusingMavenSettingupadevelopmentenvironmentusingEclipseSettingupMahoutforaWindowsuser

Beforeanyofthesteps,someoftheprerequisitesare:

YoushouldhaveJavainstalledonyoursystem.Wikihowisagoodsourceforthisathttp://www.wikihow.com/Install-Java-on-LinuxYoushouldhaveHadoopinstalledonyoursystemfromthehttp://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-common/SingleNodeSetup.htmlURL

http://www.wikihow.com/Install-Java-on-Linux

http://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-common/SingleNodeSetup.html

BuildingMahoutfromsourceusingMavenMahout’sbuildandreleasesystemisbasedonMaven.

InstallingMaven1. Createthefolder/usr/local/maven,asfollows:

mkdir/usr/local/maven

2. Downloadthedistributionapache-maven-x.y.z-bin.tar.gzfromtheMavensite(http://maven.apache.org/download.cgi)andmovethisto/usr/local/maven,asfollows:

mvapache-maven-x.y.z-bin.tar.gz/usr/local/maven

3. Unpacktothelocation/usr/local/maven,asfollows:

tar–xvfapache-maven-x.y.z-bin.tar.gz

4. Editthe.bashrcfile,asfollows:

exportM2_HOME=/usr/local/apache-maven-x.y.z

exportM2=$M2_HOME/bin

exportPATH=$M2:$PATH

NoteFortheEclipseIDE,gotoHelpandselectInstallnewSoftware.ClickontheAddbutton,andinthepopup,typethenameM2Eclipse,providethelinkhttp://download.eclipse.org/technology/m2e/releases,andclickonOK.

BuildingMahoutcodeBydefault,MahoutassumesthatHadoopisalreadyinstalledonthesystem.MahoutusestheHADOOP_HOMEandHADOOP_CONF_DIRenvironmentvariablestoaccessHadoopclusterconfigurations.ForsettingupMahout,executethefollowingsteps:

1. DownloadtheMahoutdistributionfilemahout-distribution-0.9-src.tar.gzfromthelocationhttp://archive.apache.org/dist/mahout/0.9/.

2. ChooseaninstallationdirectoryforMahout(/usr/local/Mahout),andplacethedownloadedsourceinthefolder.Extractthesourcecodeandensurethatthefoldercontainsthepom.xmlfile.Thefollowingistheexactlocationofthesource:

tar-xvfmahout-distribution-0.9-src.tar.gz

3. InstalltheMahoutMavenproject,andskipthetestcaseswhileinstalling,asfollows:

mvninstall-Dmaven.test.skip=true

4. SettheMAHOUT_HOMEenvironmentvariableinthe~/.bashrcfile,andupdatethePATHvariablewiththeMahoutbindirectory:

exportMAHOUT_HOME=/user/local/mahout/mahout-distribution-0.9

http://maven.apache.org/download.cgi

http://download.eclipse.org/technology/m2e/releases

http://archive.apache.org/dist/mahout/0.9/

exportPATH=$PATH:$MAHOUT_HOME/bin

5. TotesttheMahoutinstallation,executethecommand:mahout.Thiswilllisttheavailableprogramswithinthedistributionbundle,asshowninthefollowingscreenshot:

SettingupadevelopmentenvironmentusingEclipseForthissetup,youshouldhaveMaveninstalledonthesystemandtheMavenpluginforEclipse.RefertotheInstallingMavenstepexplainedintheprevioussection.Thissetupcanbedoneinthefollowingsteps:

1. DownloadtheMahoutdistributionfilemahout-distribution-0.9-src.tar.gzfromthelocationhttp://archive.apache.org/dist/mahout/0.9/andunzipthis:

tarxzfmahout-distribution-0.9-src.tar.gz

2. Let’screateafoldernamedworkspaceunder/usr/local/workspace,asfollows:

mkdir/usr/local/workspace

3. Movethedownloadeddistributiontothisfolder(fromthedownloadsfolder),asfollows:

mvmahout-distribution-0.9/usr/local/workspace/

4. Movetothefolder/usr/local/workspace/mahout-distribution-0.9andmakeanEclipseproject(thiscommandcantakeuptoanhour):

mvneclipse:eclipse

5. SettheMahouthomeinthe.bashrcfile,asexplainedearlierintheBuildingMahoutcodesection.

6. NowopenEclipse.Selectthefile,importMaven,andExistingMavenProjects.Now,navigatetothelocationformahout-distribution-0.9andclickonFinish.

http://archive.apache.org/dist/mahout/0.9/

SettingupMahoutforaWindowsuserAWindowsusercanuseCygwin(alargecollectionofGNUandopensourcetoolsthatprovidesfunctionalitysimilartoaLinuxdistributiononWindows)tosetuptheirenvironment.Thereisalsoanotherwaythatiseasytouse,asshowninthefollowingsteps:

1. DownloadHortonworksSandboxforvirtualboxonyoursystemfromthelocationhttp://hortonworks.com/products/hortonworks-sandbox/#install.HortonworksSandboxonyoursystemwillbeapseudo-distributedmodeofHadoop.

2. Logintotheconsole.UseAlt+F5oralternativelydownloadPuttyandprovide127.0.0.1asthehostnameand2222intheport,asshowninthefollowingfigure.Loginwiththeusernamerootandpassword-hadoop.

3. Enterthefollowingcommand:

yuminstallmahout

Now,youwillseeascreenlikethis:

http://hortonworks.com/products/hortonworks-sandbox/#install

4. Entery,andyourMahoutwillstartinstalling.Oncethisisdone,youcantestbytypingthecommandmahoutandthiswillshowyouthesamescreenasshownintheSettingupadevelopmentenvironmentusingEclipserecipeseenearlier.

SummaryWediscussedApacheMahoutindetailinthischapter.WecoveredtheprocessofinstallingMahoutonoursystem,alongwithsettingupadevelopmentenvironmentthatisreadytoexecuteMahoutalgorithms.WehavealsotakenalookatthereasonsbehindMahoutbeingconsideredagoodchoiceforclassification.Now,wemovetothenextwherewewillunderstandaboutlogisticregressionandlearnabouttheprocessthatneedstobefollowedtoexecuteourfirstalgorithminMahout.

Chapter3.LearningLogisticRegression/SGDUsingMahoutInsteadofjumpingdirectlyintologisticregression,let’strytounderstandafewofitsconcepts.Inthischapter,wewillexplorethefollowingtopics:

IntroducingregressionUnderstandinglinearregressionCostfunctionGradientdescentLogisticregressionUnderstandingSGDUsingMahoutforlogisticregression

IntroducingregressionRegressionanalysisisusedforpredictionandforecasting.Itisusedtofindouttherelationshipbetweenexplanatoryvariablesandtargetvariables.Essentially,itisastatisticalmodelthatisusedtofindouttherelationshipamongvariablespresentinthedatasets.Anexamplethatyoucanrefertoforabetterunderstandingofthistermisthis:determinetheearningsofworkersinaparticularindustry.Here,wewilltrytofindoutthefactorsthataffectaworker’ssalary.Thesefactorscanbeage,education,yearsofexperience,particularskillset,location,andsoon.Wewilltrytomakeamodelthatwilltakeallthesevariablesintoconsiderationandtrytopredictthesalary.Inregressionanalysis,wecharacterizethevariationofthetargetvariablearoundtheregressionfunction,whichcanbedescribedbyaprobabilitydistributionthatisalsoofinterest.Thereareanumberofregressionanalysistechniquesthatareavailable.Forexample,linearregression,ordinaryleastsquaresregression,logisticregression,andsoon.

UnderstandinglinearregressionInlinearregression,wecreateamodeltopredictthevalueofatargetvariablewiththehelpofanexplanatoryvariable.Tounderstandthisbetter,let’slookatanexample.

AcompanyXthatdealsinsellingcoffeehasnoticedthatinthemonthofmonsoon,theirsalesincreasedtoquiteanextent.Sotheyhavecomeupwithaformulatofindtherelationbetweenrainandtheirpercupcoffeesale,whichisshownasfollows:

C=1.5R+800

So,for2mmofrain,thereisademandof803cupsofcoffee.Nowifyougointominutedetails,youwillrealizethatwehavethedataforrainfallandpercupcoffeesale,andwearetryingtobuildamodelthatcanpredictthedemandforcoffeebasedontherainfall.Wehavedataintheformof(R1,C1),(R2,C2)….(Ri,Ci).Here,wewillbuildthemodelinamannerthatkeepstheerrorintheactualandpredictedvaluesataminimum.

CostfunctionIntheequationC=1.5R+800,thetwovalues1.5and800areparametersandthesevaluesaffecttheendresult.WecanwritethisequationasC=p0+p1R.Aswediscussedearlier,ourgoalistoreducethedifferencebetweentheactualvalueandthepredictedvalue,andthisisdependentonthevaluesofp0andp1.Let’sassumethatthepredictedvalueisCpandtheactualvalueisCsothatthedifferencewillbe(Cp-C).Thiscanbewrittenas(p0+p1R-C).Tominimizethiserror,wedefinetheerrorfunction,whichisalsocalledthecostfunction.

Thecostfunctioncanbedefinedwiththefollowingformula:

Here,iistheithsampleandNisthenumberoftrainingexamples.Wecalculatecostsfordifferentsetsofp0andp1andfinallyselectthep0andp1thatgivestheleastcost(C).Thisisthemodelthatwillbeusedtomakepredictionsfornewinput.

GradientdescentGradientdescentstartswithaninitialsetofparametervalues,p0andp1,anditerativelymovestowardsasetofparametervaluesthatminimizesthecostfunction.Wecanvisualizethiserrorfunctiongraphically,wherewidthandlengthcanbeconsideredastheparametersp0andp1andheightasthecostfunction.Ourgoalistofindthevaluesforp0andp1inawaythatourcostfunctionwillbeminimal.Westartthealgorithmwithsomevaluesofp0andp1anditerativelyworktowardstheminimumvalue.Agoodwaytoensurethatthegradientdescentisworkingcorrectlyistomakesurethatthecostfunctiondecreasesforeachiteration.Inthiscase,thecostfunctionsurfaceisconvexandwewilltrytofindouttheminimumvalue.Thiscanbeseeninthefollowingfigure:

LogisticregressionLogisticregressionisusedtoascertaintheprobabilityofanevent.Generally,logisticregressionreferstoproblemswheretheoutcomeisbinary,forexample,inbuildingamodelthatisbasedonacustomer’sincome,traveluses,gender,andotherfeaturestopredictwhetherheorshewillbuyaparticularcarornot.So,theanswerwillbeasimpleyesorno.Whentheoutcomeiscomposedofmorethanonecategory,thisiscalledmultinomiallogisticregression.

Logisticregressionisbasedonthesigmoidfunction.Predictorvariablesarecombinedwithlinearweightandthenpassedtothisfunction,whichgeneratestheoutputintherangeof0–1.Anoutputcloseto1indicatesthatanitembelongstoacertainclass.Let’sfirstunderstandthesigmoidorlogisticfunction.Itcanbedefinedbythefollowingformula:

F(z)=1/1+e(-z)

Withasingleexplanatoryvariable,zwillbedefinedasz=β0+β1*x.Thisequationisexplainedasfollows:

z:Thisiscalledthedependentvariable.Thisisthevariablethatwewouldliketopredict.Duringthecreationofthemodel,wehavethisvariablewithusinthetrainingset,andwebuildthemodeltopredictthisvariable.Theknownvaluesofzarecalledobservedvalues.x:Thisistheexplanatoryorindependentvariable.Thesevariablesareusedtopredictthedependentvariablez.Forexample,topredictthesalesofanewlylaunchedproductataparticularlocation,wemightincludeexplanatoryvariablessuchasthepriceoftheproduct,theaverageincomeofthepeopleofthatlocation,andsoon.β0:Thisiscalledtheregressionintercept.Ifallexplanatoryvariablesarezero,thenthisparameterisequaltothedependentvariablez.β1:Thesearevaluesforeachexplanatoryvariable.

Thegraphofthelogisticfunctionisasfollows:

Withalittlebitofmathematics,wecanchangethisequationasfollows:

ln(F(x)/(1-F(x))=β0+β1*x

Inthecaseoflinearregression,thecostfunctiongraphwasconvex,buthere,itisnotgoingtobeconvex.Findingtheminimumvaluesforparametersinawaythatourpredictedoutputisclosetotheactualonewillbedifficult.Inacostfunction,whilecalculatingforlogisticregression,wewillreplaceourCpvalueoflinearregressionwiththefunctionF(z).Tomakeconvexlogisticregressioncostfunctions,wewillreplace(p0+p1Ri-Ci)2withoneofthefollowing:

log(1/1+e(-(β0+β1*x)))iftheactualoccurrenceofaneventis1,thisfunctionwillrepresentthecost.log(1-(1/1+e(-(β0+β1*x))))iftheactualoccurrenceofaneventis0,thisfunctionwillrepresentthecost.

Wewillhavetorememberthatinlogisticregression,wecalculatetheclassprobability.So,iftheprobabilityofaneventoccurring(customerbuyingacar,beingdefrauded,andsoon)isp,theprobabilityofnon-occurrenceis1-p.

StochasticGradientDescentGradientdescentminimizesthecostfunction.Forverylargedatasets,gradientdescentisaveryexpensiveprocedure.StochasticGradientDescent(SGD)isamodificationofthegradientdescentalgorithmtohandlelargedatasets.Gradientdescentcomputesthegradientusingthewholedataset,whileSGDcomputesthegradientusingasinglesample.So,gradientdescentloadsthefulldatasetandtriestofindoutthelocalminimumonthegraphandthenrepeatthefullprocessagain,whileSGDadjuststhecostfunctionforeverysample,onebyone.AmajoradvantagethatSGDhasovergradientdescentisthatitsspeedofcomputationisawholelotfaster.LargedatasetsinRAMgenerallycannotbeheldasthestorageislimited.InSGD,theburdenontheRAMisreduced,whereineachsampleorbatchofsamplesareloadedandworkedwith,theresultsforwhicharestored,andsoon.

UsingMahoutforlogisticregressionMahouthasimplementationsforlogisticregressionusingSGD.Itisveryeasytounderstandanduse.Solet’sgetstarted.

Dataset

WewillusetheWisconsinDiagnosticBreastCancer(WDBC)dataset.Thisisadatasetforbreastcancertumorsanddataisavailablefrom1995onwards.Ithas569instancesofbreasttumorcasesandhas30featurestopredictthediagnosis,whichiscategorizedaseitherbenignormalignant.

NoteMoredetailsontheprecedingdatasetisavailableathttp://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.names.

Preparingthetrainingandtestdata

Youcandownloadthewdbc.datadatasetfromhttp://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.data.

Now,saveitasaCSVfileandincludethefollowingheaderline:ID_Number,Diagnosis,Radius,Texture,Perimeter,Area,Smoothness,Compactness,Concavity,ConcavePoints,Symmetry,Fractal_Dimension,RadiusStdError,TextureStdError,PerimeterStdError,AreaStdError,SmoothnessStdError,CompactnessStdError,ConcavityStdError,ConcavePointStdError,Symmetrystderror,FractalDimensionStderror,WorstRadius,worsttexture,worstperimeter,worstarea,worstsmoothness,worstcompactness,worstconcavity,worstconcavepoints,worstsymmentry,worstfractaldimensions

Now,wewillhavetoperformthefollowingstepstopreparethisdatatobeusedbytheMahoutlogisticregressionalgorithm:

1. Wewillmakethetargetclassnumeric.Inthiscase,thesecondfielddiagnosisisthetargetvariable.Wewillchangemalignantto0andbenignto1.Usethefollowingcodesnippettointroducethechanges.Wecanusethisstrategyforsmalldatasets,butforhugedatasets,wehavedifferentstrategies,whichwewillcoverinChapter4,LearningtheNaïveBayesClassificationUsingMahout:

publicvoidconvertTargetToInteger()throwsIOException{

//Readthedata

BufferedReaderbr=newBufferedReader(newFileReader("wdbc.csv"));

Stringline=null;

//Createthefiletosavetheresulteddata

FilewdbcData=newFile("<YourDestinationlocationforfile.>");

FileWriterfw=newFileWriter(wdbcData);

//Weareaddingheadertothenewfile

fw.write("ID_Number"+","+"Diagnosis"+","+"Radius"+","+"Texture"+","+"Pe

rimeter"+","+"Area"+","+"Smoothness"+","+"Compactness"+","+"Concavity"+

","+"ConcavePoints"+","+"Symmetry"+","+"Fractal_Dimension"+","+"RadiusS

tdError"+","+"TextureStdError"+","+"PerimeterStdError"+","+"AreaStdErro

r"+","+"SmoothnessStdError"+","+"CompactnessStdError"+","+"ConcavityStd

Error"+","+"ConcavePointStdError"+","+"Symmetrystderror"+","+"FractalDi

mensionStderror"+","+"WorstRadius"+","+"worsttexture"+","+"worstperimet

er"+","+"worstarea"+","+"worstsmoothness"+","+"worstcompactness"+","+"w

orstconcavity"+","+"worstconcavepoints"+","+"worstsymmentry"+","+"worst

fractaldimensions"+"\n");

http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.names

http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/wdbc.data

/*Inthewhileloopwearereadinglinebylineandcheckingthelast

field-parts[1]andchangingittonumericvalueaccordingly*/

while((line=br.readLine())!=null){

String[]parts=line.split(",");

if(parts[1].equals("M")){

fw.write(parts[0]+","+"0"+","+parts[2]+","+parts[3]+","+parts[4]+","+pa

rts[5]+","+parts[6]+","+parts[7]+","+parts[8]+","+parts[9]+","+parts[10

]+","+parts[11]+","+parts[12]+","+parts[13]+","+parts[14]+","+parts[15]

+","+parts[16]+","+parts[17]+","+parts[18]+","+parts[19]+","+parts[20]+

","+parts[21]+","+parts[22]+","+parts[23]+","+parts[24]+","+parts[25]+"

,"+parts[26]+","+parts[27]+","+parts[28]+","+parts[29]+","+parts[30]+",

"+parts[31]+"\n");

}

if(parts[1].equals("B")){

fw.write(parts[0]+","+"1"+","+parts[2]+","+parts[3]+","+parts[4]+","+pa

rts[5]+","+parts[6]+","+parts[7]+","+parts[8]+","+parts[9]+","+parts[10

]+","+parts[11]+","+parts[12]+","+parts[13]+","+parts[14]+","+parts[15]

+","+parts[16]+","+parts[17]+","+parts[18]+","+parts[19]+","+parts[20]+

","+parts[21]+","+parts[22]+","+parts[23]+","+parts[24]+","+parts[25]+"

,"+parts[26]+","+parts[27]+","+parts[28]+","+parts[29]+","+parts[30]+",

"+parts[31]+"\n");

}

}

fw.close();

br.close();

}

TipDownloadingtheexamplecode

Youcandownloadtheexamplecodefilesfromyouraccountathttp://www.packtpub.comforallthePacktPublishingbooksyouhavepurchased.Ifyoupurchasedthisbookelsewhere,youcanvisithttp://www.packtpub.com/supportandregistertohavethefilese-maileddirectlytoyou.

2. Wewillhavetosplitthedatasetintotrainingandtestdatasetsandthenshufflethedatasetssothatwecanmixthemup,whichcanbedoneusingthefollowingcodesnippet:

publicvoiddataPrepration()throwsException{

//Readingthedatasetcreatedbyearliermethod

convertTargetToIntegerandhereweareusinggoogleguavaapi's.

List<String>result=

Resources.readLines(Resources.getResource("wdbc.csv"),Charsets.UTF_8);

//Thisistoremoveheaderbeforetherandomizationprocess.

Otherwiseitcanappearinthemiddleofdataset.

List<String>raw=result.subList(1,570);

Randomrandom=newRandom();

//Shufflingthedataset.

Collections.shuffle(raw,random);

//Splittingdatasetintotrainingandtestexamples.


http://www.packtpub.com/support

List<String>train=raw.subList(0,470);

List<String>test=raw.subList(470,569);

FiletrainingData=newFile("<yourLocation>/wdbcTrain.csv");

FiletestData=newFile("<yourLocation>/wdbcTest.csv");

writeCSV(train,trainingData);

writeCSV(test,testData);

}

//Thismethodiswritingthelisttodesiredfilelocation.

publicvoidwriteCSV(List<String>list,Filefile)throwsIOException{

FileWriterfw=newFileWriter(file);

fw.write("ID_Number"+","+"Diagnosis"+","+"Radius"+","+"Texture"+","+"Pe

rimeter"+","+"Area"+","+"Smoothness"+","+"Compactness"+","+"Concavity"+

","+"ConcavePoints"+","+"Symmetry"+","+"Fractal_Dimension"+","+"RadiusS

tdError"+","+"TextureStdError"+","+"PerimeterStdError"+","+"AreaStdErro

r"+","+"SmoothnessStdError"+","+"CompactnessStdError"+","+"ConcavityStd

Error"+","+"ConcavePointStdError"+","+"Symmetrystderror"+","+"FractalDi

mensionStderror"+","+"WorstRadius"+","+"worsttexture"+","+"worstperimet

er"+","+"worstarea"+","+"worstsmoothness"+","+"worstcompactness"+","+"w

orstconcavity"+","+"worstconcavepoints"+","+"worstsymmentry"+","+"worst

fractaldimensions"+"\n");

for(inti=0;i<list.size();i++){

fw.write(list.get(i)+"\n");

}

fw.close();

}

Trainingthemodel

Wewillusethetrainingdatasetandtrainlogisticalgorithmtopreparethemodel.Usethefollowingcommandtocreatethemodel:

mahouttrainlogistic--input/tmp/wdbcTrain.csv--output/tmp//model--

targetDiagnosis--categories2--predictorsRadiusTexturePerimeterArea

SmoothnessCompactnessConcavityConcavePointsSymmetryFractal_Dimension

RadiusStdErrorTextureStdErrorPerimeterStdErrorAreaStdError

SmoothnessStdErrorCompactnessStdErrorConcavityStdError

ConcavePointStdErrorSymmetrystderrorFractalDimensionStderrorWorstRadius

worsttextureworstperimeterworstareaworstsmoothnessworstcompactness

worstconcavityworstconcavepointsworstsymmentryworstfractaldimensions--

typesnumeric--features30--passes90--rate300

Thiscommandwillgiveyouthefollowingoutput:

Let’sunderstandtheparametersusedinthiscommand:

trainlogistic:ThisisthealgorithmthatMahoutprovidestobuildthemodelusingyourinputparameters.input:Thisisthelocationoftheinputfile.output:Thisisthelocationofthemodelfile.target:Thisisthenameofthetargetvariablethatwewanttopredictfromthedataset.categories:Thisreferstothenumberofpredictedclasses.predictors:Thisfeaturesinthedatasetusedtopredictthetargetvariable.types:Thisisalistofthetypesofpredictorvariables.(Hereallarenumericbutitcouldbewordortextaswell.)features:Thisisthesizeofthefeaturevectorusedtobuildthemodel.passes:Thisspecifiesthenumberoftimestheinputdatashouldbere-examinedduringtraining.Smallinputfilesmayneedtobeexamineddozensoftimes.Verylargeinputfilesprobablydon’tevenneedtobecompletelyexamined.rate:Thissetstheinitiallearningrate.Thiscanbelargeifyouhavelotsofdataoruselotsofpassesbecauseitdecreasesprogressivelyasdataisexamined.

Nowourmodelisreadytomoveontothenextstepofevaluation.Toevaluatethemodelfurther,wecanusethesamedatasetandchecktheconfusionandAUCmatrix.Thecommandforthiswillbeasfollows:

mahoutrunlogistic--input/tmp/wdbcTrain.csv--model/tmp//model--auc--

confusion

runlogistic:Thisisthealgorithmtorunthelogisticregressionmodeloveraninputdatasetmodel:Thisisthelocationofthemodelfileauc:ThisprintstheAUCscoreforthemodelversustheinputdataafterthedataisreadconfusion:Thisprintstheconfusionmatrixforaparticularthreshold

Theoutputofthepreviouscommandisshowninthefollowingscreenshot:

Now,thesematricesshowthatthemodelisnotbad.Having0.88asthevalueforAUCisgood,butwewillcheckthisontestdataaswell.Theconfusionmatrixinformsusthatoutof172malignanttumors,ithascorrectlyclassified151instancesandthat34benigntumorsarealsoclassifiedasmalignant.Inthecaseofbenigntumors,outof298,ithascorrectlyclassified264.

Ifthemodeldoesnotprovidegoodresults,wehaveanumberofoptions.

Changetheparametersinthefeaturevector,increasingthemifweareselectingfewfeatures.Thisshouldbedoneoneatatime,andweshouldtesttheresultagainwitheachgeneratedmodel.WeshouldgetamodelwhereAUCiscloseto1.

Let’srunthesamealgorithmontestdataaswell:

mahoutrunlogistic--input/tmp/wdbcTest.csv--model/tmp//model--auc–

confusion

Sothismodelworksalmostthesameontestdataaswell.Ithasclassified34outofthe40malignanttumorscorrectly.

SummaryInthischapter,wediscussedlogisticregressionandhowwecanusethisalgorithmavailableinApacheMahout.WeusedtheWisconsinDiagnosticBreastCancerdatasetandrandomlybrokeitintotwodatasets:onefortrainingandtheotherfortesting.WecreatedthelogisticregressionmodelusingMahoutandalsorantestdataoverthismodel.Now,wewillmoveontothenextchapterwhereyouwilllearnabouttheNaïveBayesclassificationandalsothemostfrequentlyusedclassificationtechnique:textclassification.

Chapter4.LearningtheNaïveBayesClassificationUsingMahoutInthischapter,wewillusetheNaïveBayesclassificationalgorithmtoclassifyasetofdocuments.Classifyingtextdocumentsisalittletrickybecauseofthedatapreparationstepsinvolved.Inthischapter,wewillexplorethefollowingtopics:

ConditionalprobabilityandtheBayesruleUnderstandingtheNaïveBayesalgorithmUnderstandingtermsusedintextclassificationUsingtheNaïveBayesalgorithminApacheMahout

IntroducingconditionalprobabilityandtheBayesruleBeforelearningtheNaïveBayesalgorithm,youshouldhaveanunderstandingofconditionalprobabilityandtheBayesrule.

Inverysimpleterms,conditionalprobabilityistheprobabilitythatsomethingwillhappen,giventhatsomethingelsehasalreadyhappened.ItisexpressedasP(A/B),whichcanbereadasprobabilityofAgivenB,anditfindstheprobabilityoftheoccurrenceofeventAonceeventBhasalreadyhappened.

Mathematically,itisdefinedasfollows:

Forexample,ifyouchooseacardfromastandardcarddeckandifyouwereaskedabouttheprobabilityforthecardtobeadiamond,youwouldquicklysay13/52or0.25,asthereare13diamondcardsinthedeck.However,ifyouthenlookatthecardanddeclarethatitisred,thenwewillhavenarrowedthepossibilitiesforthecardto26possiblecards,andtheprobabilitythatthecardisadiamondnowis13/26=0.5.So,ifwedefineAasadiamondcardandBasaredcard,thenP(A/B)willbetheprobabilityofthecardbeingadiamond,givenitisred.

Sometimes,foragivenpairofevents,conditionalprobabilityishardtocalculate,andBayes’theoremhelpsusherebygivingtherelationshipbetweentwoconditionalprobabilities.

Bayes’theoremisdefinedasfollows:

Thetermsintheformulaaredefinedasfollows:

P(A):ThisiscalledpriorprobabilityorpriorP(B/A):ThisiscalledconditionalprobabilityorlikelihoodP(B):ThisiscalledmarginalprobabilityP(A/B):Thisiscalledposteriorprobabilityorposterior

Thefollowingformulaisderivedonlyfromtheconditionalprobabilityformula.WecandefineP(B/A)asfollows:

Whenrearranged,theformulabecomesthis:

Now,fromtheprecedingconditionalprobabilityformula,wegetthefollowing:

Let’stakeanexamplethatwillhelpustounderstandhowBayes’theoremisapplied.

Acancertestgivesapositiveresultwithaprobabilityof97percentwhenthepatientisindeedaffectedbycancer,whileitgivesanegativeresultwith99percentprobabilitywhenthepatientisnotaffectedbycancer.Ifapatientisdrawnatrandomfromapopulationwhere0.2percentoftheindividualsareaffectedbycancerandheorsheisfoundtobepositive,whatistheprobabilitythatheorsheisindeedaffectedbycancer?Inprobabilisticterms,whatweknowaboutthisproblemcanbedefinedasfollows:

P(positive|cancer)=0.97

P(positive|nocancer)=1-0.99=0.01

P(cancer)=0.002

P(nocancer)=1-0.002=0.998

P(positive)=P(positive|cancer)P(cancer)+P(positive|nocancer)P(nocancer)

=0.97*0.002+0.01*0.998

=0.01192

NowP(cancer|positive)=(0.97*0.002)/0.01192=0.1628

Soevenwhenfoundpositive,theprobabilityofthepatientbeingaffectedbycancerinthisexampleisaround16percent.

UnderstandingtheNaïveBayesalgorithmInBayes’theorem,wehaveseenthattheoutcomeisbasedonlyononeevidence,butinclassificationproblems,wehavemultipleevidencesandwehavetopredicttheoutcome.InNaïveBayes,weuncouplemultiplepiecesofevidenceandtreateachoneofthemindependently.Itisdefinedasfollows:

P(outcome|multipleEvidence))=P(Evidence1|outcome)*P(Evidence2|outcome)*P(Evidence3|outcome)…./P(Evidence)

Runthisformulaforeachpossibleoutcome.Sincewearetryingtoclassify,eachoutcomewillbecalledaclass.Ourtaskistolookattheevidence(features)toconsiderhowlikelyitisforittobeofaparticularclassandthenassignitaccordingly.Theclassthathasthehighestprobabilitygetsassignedtothatcombinationofevidences.Let’sunderstandthiswithanexample.

Let’ssaythatwehavedataon1,000piecesoffruit.Theyhappentobebananas,apples,orsomeotherfruit.Weareawareofthreecharacteristicsofeachfruit:

Size:TheyareeitherlongornotlongTaste:TheyareeithersweetornotsweetColor:Theyareeitheryellowornotyellow

Assumethatwehaveadatasetlikethefollowing:

Fruittype Taste–sweet Taste–notsweet Color–yellow Color–notyellow Size–long Size–notlong Total

Banana 350 150 450 50 400 100 500

Apple 150 150 100 200 0 300 300

Other 150 50 50 150 100 100 200

Total 650 350 600 400 500 500 1000

Nowlet’slookatthethingswehave:

P(Banana)=500/1000=0.5

P(Apple)=300/1000=0.3

P(Other)=200/1000=0.2

Let’slookattheprobabilityofthefeatures:

P(Sweet)=650/1000=0.65

P(Yellow)=600/1000=0.6

P(long)=500/1000=0.5

P(notSweet)=350/1000=0.35

P(notyellow)=400/1000=0.4

P(notlong)=500/1000=0.5

Nowwewanttoknowwhatfruitwewillhaveifitisnotyellowandnotlongandsweet.Theprobabilityofitbeinganappleisasfollows:

P(Apple|sweet,notlong,notyellow)=P(sweet|Apple)*P(notlong|Apple)*P(notyellow|Apple)*P(Apple)/P(sweet)*P(notlong)*P(notyellow)

=0.5*1*0.67*0.3/P(Evidence)

=0.1005/P(Evidence)

Theprobabilityofitbeingabananaisthis:

P(banana|sweet,notlong,notyellow)=P(sweet|banana)*P(notlong|banana)*P(notyellow|banana)*P(banana)/P(sweet)*P(notlong)*P(notyellow)

=0.7*0.2*0.1*0.5/P(Evidence)

=0.007/P(Evidence)

Theprobabilityofitbeinganyotherfruitisasfollows:

P(otherfruit|sweet,notlong,notyellow)=P(sweet|otherfruit)*P(notlong|otherfruit)*P(notyellow|otherfruit)*P(otherfruit)/P(sweet)*P(notlong)*P(notyellow)

=0.75*0.5*0.75*0.2/P(Evidence)

=0.05625/P(Evidence)

Sofromtheresults,youcanseethatifthefruitissweet,notlong,andnotyellow,thenthehighestprobabilityisthatitwillbeanapple.Sofindoutthehighestprobabilityandassigntheunknownitemtothatclass.

NaïveBayesisaverygoodchoicefortextclassification.BeforewemoveontotextclassificationusingNaïveBayesinMahout,let’sunderstandafewtermsthatarereallyusefulfortextclassification.

UnderstandingthetermsusedintextclassificationTopreparedatasothatitcanbeusedbyaclassifierisacomplexprocess.Fromrawdata,wecancollectexplanatoryandtargetvariablesandencodethemasvectors,whichistheinputoftheclassifier.

Vectorsareorderedlistsofvaluesasdefinedintwo-dimensionalspace.Youcantakeacluefromcoordinategeometryaswell.Apoint(3,4)isapointinthexandyplanes.InMahout,itisdifferent.Here,avectorcanhave(3,4)or10,000dimensions.

Mahoutprovidessupportforcreatingvectors.TherearetwotypesofvectorimplementationsinMahout:sparseanddensevectors.Thereareafewtermsthatweneedtounderstandfortextclassification:

Bagofwords:Thisconsiderseachdocumentasacollectionofwords.Thisignoreswordorder,grammar,andpunctuation.So,ifeverywordisafeature,thencalculatingthefeaturevalueofthedocumentwordisrepresentedasatoken.Itisgiventhevalue1ifitispresentor0ifnot.Termfrequency:Thisconsidersthewordcountinthedocumentinsteadof0and1.Sotheimportanceofawordincreaseswiththenumberoftimesitappearsinthedocument.Considerthefollowingexamplesentence:

ApplehaslaunchediPhoneanditwillcontinuetolaunchsuchproducts.OthercompetitorsarealsoplanningtolaunchproductssimilartothatofiPhone.

Thefollowingisthetablethatrepresentstermfrequency:

Term Count

Apple 1

Launch 3

iPhone 2

Product 2

Plan 1

Thefollowingtechniquesareusuallyappliedtocomeupwiththistypeoftable:

Stemmingofwords:Withthis,thesuffixisremovedfromthewordso“launched”,“launches”,and“launch”areallconsideredas“launch”.Casenormalization:Withthis,everytermisconvertedtolowercase.Stopwordremoval:Therearesomewordsthatarealmostpresentineverydocument.Wecallthesewordsstopwords.Duringanimportantfeatureextractionfromadocument,thesewordscomeintoaccountandtheywillnotbehelpfulinthe

overallcalculation.Examplesofthesewordsare“is,are,the,that,andsoon.”So,whileextracting,wewillignorethesekindofwords.Inversedocumentfrequency:Thisisconsideredastheboostatermgetsforbeingrare.Atermshouldnotbetoocommon.Ifatermoccursineverydocument,itisnotgoodforclassification.Thefewerdocumentsinwhichatermoccurs,themoresignificantitislikelytobeforthedocumentsitdoesoccurin.Foratermt,inversedocumentfrequencyiscalculatedasfollows:

IDF(t)=1+log(totalnumberofdocuments/numberofdocumentscontainingt)

Termfrequencyandinversetermfrequency:Thisisoneofthepopularrepresentationsofthetext.Itistheproductoftermfrequencyandinversedocumentfrequency,asfollows:

TFIDF(t,d)=TF(t,d)*IDF(t)

Eachdocumentisafeaturevectorandacollectionofdocumentsisasetofthesefeaturevectorsandthissetworksastheinputfortheclassification.Nowthatweunderstandthebasicconceptsbehindthevectorcreationoftextdocuments,let’smoveontothenextsectionwherewewillclassifytextdocumentsusingtheNaïveBayesalgorithm.

UsingtheNaïveBayesalgorithminApacheMahoutWewilluseadatasetof20newsgroupsforthisexercise.The20newsgroupsdatasetisastandarddatasetcommonlyusedformachinelearningresearch.Thedataisobtainedfromtranscriptsofseveralmonthsofpostingsmadein20Usenetnewsgroupsfromtheearly1990s.Thisdatasetconsistsofmessages,oneperfile.Eachfilebeginswithheaderlinesthatspecifythingssuchaswhosentthemessage,howlongitis,whatkindofsoftwarewasused,andthesubject.Ablanklinefollowsandthenthemessagebodyfollowsasunformattedtext.

Downloadthe20news-bydate.tar.gzdatasetfromhttp://qwone.com/~jason/20Newsgroups/.ThefollowingstepsareusedtobuildtheNaïveBayesclassifierusingMahout:

1. Createa20newsdatadirectoryandunzipthedatahere:

mkdir/tmp/20newsdata

cd/tmp/20newsdata

tar–xzvf/tmp/20news-bydate.tar.gz

2. Youwillseetwofoldersunder20newsdata:20news-bydate-testand20news-bydate-train.Nowcreateanotherdirectorycalled20newsdataallandmergeboththetrainingandtestdataofthe20newsgroups.

3. Comeoutofthedirectoryandmovetothehomedirectoryandexecutethefollowing:

mkdir/tmp/20newsdataall

cp–R/20newsdata/*/*/tmp/20newsdataall

4. CreateadirectoryinHadoopandsavethisdatainHDFSformat:

hadoopfs–mkdir/user/hue/20newsdata

hadoopfs–put/tmp/20newsdataall/user/hue/20newsdata

5. Converttherawdataintoasequencefile.Theseqdirectorycommandwillgeneratesequencefilesfromadirectory.SequencefilesareusedinHadoop.Asequencefileisaflatfilethatconsistsofbinarykey/valuepairs.WeareconvertingthefilesintosequencefilessothatitcanbeprocessedinHadoop,whichcanbedoneusingthefollowingcommand:

bin/mahoutseqdirectory-i/user/hue/20newsdata/20newsdataall-o

/user/hue/20newsdataseq-out

Theoutputoftheprecedingcommandcanbeseeninthefollowingscreenshot:

http://qwone.com/~jason/20Newsgroups/

6. Convertthesequencefileintoasparsevectorusingthefollowingcommand:

bin/mahoutseq2sparse-i/user/hue/20newsdataseq-out/part-m-00000-o

/user/hue/20newsdatavec-lnorm-nv-wttfidf

Thetermsusedintheprecedingcommandareasfollows:

lnorm:Thisisfortheoutputvectortobelognormalizednv:Thisreferstonamedvectorswt:Thisreferstothekindofweighttouse;here,weusetfidf

Theoutputoftheprecedingcommandontheconsoleisshowninthefollowingscreenshot:

7. Splitthesetofvectorstotrainandtestthemodel:

bin/mahoutsplit-i/user/hue/20newsdatavec/tfidf-vectors--

trainingOutput/user/hue/20newsdatatrain--testOutput

/user/hue/20newsdatatest--randomSelectionPct40--overwrite--

sequenceFiles-xmsequential

Thetermsusedintheprecedingcommandareasfollows:

randomSelectionPct:Thisdividesthepercentageofdataintotestingandtrainingdatasets.Here,60percentisfortestingand40percentfortraining.xm:Thisreferstotheexecutionmethodtouse:sequentialormapreduce.Thedefaultismapreduce.

8. Nowtrainthemodel:

bin/mahouttrainnb-i/user/hue/20newsdatatrain-el-o/user/hue/model

-li/user/hue/labelindex-ow-c

9. Testthemodelusingthefollowingcommand:

bin/mahouttestnb-i/user/hue/20newsdatatest-m/user/hue/model/-l

/user/hue/labelindex-ow-o/user/hue/results

Theoutputoftheprecedingcommandontheconsoleisshowninthefollowingscreenshot:

WegettheresultofourNaïveBayesclassifierforthe20newsgroups.

SummaryInthischapter,wediscussedtheNaïveBayesalgorithm.Thisalgorithmisasimplisticyethighlyregardedstatisticalmodelthatiswidelyusedinbothindustryandacademia,anditproducesgoodresultsonmanyoccasions.WeinitiallydiscussedconditionalprobabilityandtheBayesrule.WethensawanexampleoftheNaïveBayesalgorithm.Youlearnedabouttheapproachestoconverttextintoavectorformat,whichisaninputforclassifiers.Finally,weusedthe20newsgroupsdatasettobuildaclassifierusingtheNaïveBayesalgorithminMahout.Inthenextchapter,wewillcontinueourjourneyofexploringclassificationalgorithmsinMahoutwiththeHiddenMarkovmodelimplementation.

Chapter5.LearningtheHiddenMarkovModelUsingMahoutInthischapter,wewillcoveroneofthemostinterestingtopicsofclassificationtechniques:theHiddenMarkovModel(HMM).TounderstandtheHMM,wewillcoverthefollowingtopicsinthischapter:

DeterministicandnondeterministicpatternsTheMarkovprocessIntroducingtheHMMUsingMahoutfortheHMM

DeterministicandnondeterministicpatternsInadeterministicsystem,eachstateissolelydependentonthestateitwaspreviouslyin.Forexample,let’stakethecaseofasetoftrafficlights.Thesequenceoflightsisred→green→amber→red.So,hereweknowwhatstatewillfollowafterthecurrentstate.Oncethetransitionsareknown,deterministicsystemsareeasytounderstand.

Fornondeterministicpatterns,consideranexampleofapersonnamedBobwhohashissnacksat4:00P.M.everyday.Let’ssayhehasanyoneofthethreeitemsfromthemenu:icecream,juice,orcake.Wecannotsayforsurewhatitemhewillhavethenextday,evenifweknowwhathehadtoday.Thisisanexampleofanondeterministicpattern.

TheMarkovprocessIntheMarkovprocess,thenextstateisdependentonthepreviousstates.Ifweassumethatwehaveannstatesystem,thenthenextstateisdependentonthepreviousnstates.Thisprocessiscalledannmodelorder.IntheMarkovprocess,wemakethechoiceforthenextstateprobabilistically.So,consideringourpreviousexample,ifBobhadjuicetoday,hecanhavejuice,icecream,orcakethenextday.Inthesameway,wecanreachanystateinthesystemfromthepreviousstate.TheMarkovprocessisshowninthefollowingdiagram:

Ifwehavenstatesinaprocess,thenwecanreachanystatewithn2transitions.Wehaveaprobabilityofmovingtoanystate,andhence,wewillhaven2probabilitiesofdoingthis.ForaMarkovprocess,wewillhavethefollowingthreeitems:

States:Thisreferstothestatesinthesystem.Inourexample,let’ssaytherearethreestates:state1,state2,andstate3.Transitionmatrix:Thiswillhavetheprobabilitiesofmovingfromonestatetoanyotherstate.Anexampleofthetransitionmatrixisshowninthefollowingscreenshot:

Thismatrixshowsthatifthesystemwasinstate1yesterday,thentheprobabilityofittoremaininthesamestatetodaywillbe0.1.

Initialstatevector:Thisisthevectoroftheinitialstateofthesystem.(Anyoneofthestateswillhaveaprobabilityof1andtherestwillhaveaprobabilityof0inthisvector.)

IntroducingtheHiddenMarkovModelTheHiddenMarkovModel(HMM)isaclassificationtechniquetopredictthestatesofasystembyobservingtheoutcomeswithouthavingaccesstotheactualstatesthemselves.ItisaMarkovmodelinwhichthestatesarehidden.

Let’scontinuewithBob’ssnackexamplewesawearlier.Nowassumewehaveonemoresetofeventsinplacethatisdirectlyobservable.WeknowwhatBobhaseatenforlunchandhissnacksintakeisrelatedtohislunch.So,wehaveanobservationstate,whichisBob’slunch,andhiddenstates,whicharehissnacksintake.WewanttobuildanalgorithmthatcanforecastwhatwouldbeBob’schoiceofsnackbasedonhislunch.

InadditiontothetransitionprobabilitymatrixintheHiddenMarkovModel,wehaveonemorematrixthatiscalledanemissionmatrix.Thismatrixcontainstheprobabilityoftheobservablestate,provideditisassignedahiddenstate.Theemissionmatrixisasfollows:

P(observablestate|onestate)

So,aHiddenMarkovModelhasthefollowingproperties:

Statevector:ThiscontainstheprobabilityofthehiddenmodeltobeinaparticularstateatthestartTransitionmatrix:Thishastheprobabilitiesofahiddenstate,giventheprevioushiddenstateEmissionmatrix:Giventhatthehiddenmodelisinaparticularhiddenstate,thishastheprobabilitiesofobservingaparticularobservablestateHiddenstates:ThisreferstothestatesofthesystemthatcanbedefinedbytheHiddenMarkovModelObservablestate:Thestatesthatarevisibleintheprocess

UsingtheHiddenMarkovModel,threetypesofproblemscanbesolved.ThefirsttwoarerelatedtothepatternrecognitionproblemandthethirdtypeofproblemgeneratesaHiddenMarkovModel,givenasequenceofobservations.Let’slookatthesethreetypes

ofproblems:

Evaluation:Thisisfindingouttheprobabilityofanobservedsequence,givenanHMM.FromthenumberofdifferentHMMsthatdescribedifferentsystemsandasequenceofobservations,ourgoalwillbetofindoutwhichHMMwillmostprobablygeneratetherequiredsequence.WeusetheforwardalgorithmtocalculatetheprobabilityofanobservationsequencewhenaparticularHMMisgivenandfindoutthemostprobableHMM.Decoding:Thisisfindingthemostprobablesequenceofhiddenstatesfromsomeobservations.WeusetheViterbialgorithmtodeterminethemostprobablesequenceofhiddenstateswhenyouhaveasequenceofobservationsandanHMM.Learning:LearningisgeneratingtheHMMfromasequenceofobservations.So,ifwehavesuchasequence,wemaywonderwhichisthemostlikelymodeltogeneratethissequence.Theforward-backwardalgorithmsareusefulinsolvingthisproblem.

TheHiddenMarkovModelisusedindifferentapplicationssuchasspeechrecognition,handwrittenletterrecognition,genomeanalysis,partsofspeechtagging,customerbehaviormodeling,andsoon.

UsingMahoutfortheHiddenMarkovModelApacheMahouthastheimplementationoftheHiddenMarkovModel.Itisavailableintheorg.apache.mahout.classifier.sequencelearning.hmmpackage.

Theoverallimplementationisprovidedbyeightdifferentclasses:

HMMModel:ThisisthemainclassthatdefinestheHiddenMarkovModel.HmmTrainer:ThisclasshasalgorithmsthatareusedtotraintheHiddenMarkovModel.Themainalgorithmsaresupervisedlearning,unsupervisedlearning,andunsupervisedBaum-Welch.HmmEvaluator:ThisclassprovidesdifferentmethodstoevaluateanHMMmodel.Thefollowingusecasesarecoveredinthisclass:

Generatingasequenceofoutputstatesfromamodel(prediction)Computingthelikelihoodthatagivenmodelwillgeneratethegivensequenceofoutputstates(modellikelihood)Computingthemostlikelyhiddensequenceforagivenmodelandagivenobservedsequence(decoding)

HmmAlgorithms:ThisclasscontainsimplementationsofthethreemajorHMMalgorithms:forward,backward,andViterbi.HmmUtils:ThisisautilityclassandprovidesmethodstohandleHMMmodelobjects.RandomSequenceGenerator:Thisisacommand-linetooltogenerateasequencebythegivenHMM.BaumWelchTrainer:ThisistheclasstotrainHMMfromtheconsole.ViterbiEvaluator:Thisisalsoacommand-linetoolforViterbievaluation.

Now,let’sworkwithBob’sexample.

Thefollowingisthegivenmatrixandtheinitialprobabilityvector:

Icecream Cake Juice

0.36 0.51 0.13

Thefollowingwillbethestatetransitionmatrix:

Icecream Cake Juice

Icecream 0.365 0.500 0.135

Cake 0.250 0.125 0.625

Juice 0.365 0.265 0.370

Thefollowingwillbetheemissionmatrix:

Spicyfood Normalfood Nofood

Icecream 0.1 0.2 0.7

Cake 0.5 0.25 0.25

Juice 0.80 0.10 0.10

Nowwewillexecuteacommand-line-basedexampleofthisproblem.WehavethreehiddenstatesofwhatBob’seatenforsnacks:ice-cream,cake,orjuice.Then,wehavethreeobservablestatesofwhatheishavingatlunch:spicyfood,normalfood,ornofoodatall.Now,thefollowingarethestepstoexecutefromthecommandline:

1. Createadirectorywiththenamehmm:mkdir/tmp/hmm.Gotothisdirectoryandcreatethesampleinputfileoftheobservedstates.ThiswillincludeasequenceofBob’slunchhabit:spicyfood(state0),normalfood(state1),andnofood(state2).Executethefollowingcommand:

echo"012221100212111122200000022200000

011112222202120212110001010212121211

002202110">hmm-input

2. RuntheBaumWelchalgorithmtotrainthemodelusingthefollowingcommand:

mahoutbaumwelch-i/tmp/hmm/hmm-input-o/tmp/hmm/hmm-model-nh3-no

3-e.0001-m1000

Theparametersusedintheprecedingcommandareasfollows:

i:Thisistheinputfilelocationo:Thisistheoutputlocationforthemodelnh:Thisisthenumberofhiddenstates.Inourexample,itisthree(icecream,juice,orcake)no:Thisisthenumberofobservablestates.Inourexample,itisthree(spicy,normal,ornofood)e:Thisistheepsilonnumber.Thisistheconvergencethresholdvaluem:Thisisthemaximumiterationnumber

Thefollowingscreenshotshowstheoutputonexecutingthepreviouscommand:

3. NowwehaveanHMMmodelthatcanbeusedtobuildapredictedsequence.Wewillrunthemodeltopredictthenext15statesoftheobservablesequenceusingthefollowingcommand:

mahouthmmpredict-m/tmp/hmm/hmm-model-o/tmp/hmm/hmm-predictions-l

10


m:ThisisthepathfortheHMMmodel

o:Thisistheoutputdirectorypath

l:Thisisthelengthofthegeneratedsequence

4. Toviewthepredictionforthenext10observablestates,usethefollowingcommand:

mahouthmmpredict-m/tmp/hmm/hmm-model-o/tmp/hmm/hmm-predictions-l

10


Fromtheoutput,wecansaythatthenextobservablestatesforBob’slunchwillbespicy,spicy,spicy,normal,normal,nofood,nofood,nofood,nofood,andnofood.

5. Now,wewilluseonemorealgorithmtopredictthehiddenstate.WewillusetheViterbialgorithmtopredictthehiddenstatesforagivenobservationalstate’ssequence.Wewillfirstcreatethesequenceoftheobservationalstateusingthefollowingcommand:

echo"012021100112">/tmp/hmm/hmm-viterbi-input

6. WewillusetheViterbicommand-lineoptiontogeneratetheoutputwiththelikelihoodofgeneratingthissequence:

mahoutviterbi--input/tmp/hmm/hmm-viterbi-input--outputtmp/hmm/hmm-

viterbi-output--model/tmp/hmm/hmm-model--likelihood


input:Thisistheinputlocationofthefileoutput:ThisistheoutputlocationoftheViterbialgorithm’soutputmodel:ThisistheHMMmodellocationthatwecreatedearlierlikelihood:Thisisthecomputedlikelihoodoftheobservedsequence

Thefollowingscreenshotshowstheoutputonexecutingthepreviouscommand:

7. PredictionsfromtheViterbiaresavedintheoutputfileandcanbeseenusingthecatcommand:

cat/tmp/hmm/hmm-viterbi-output

Thefollowingoutputshowsthepredictionsforthehiddenstate:

SummaryInthischapter,wediscussedanotherclassificationtechnique:theHiddenMarkovModel.Youlearnedaboutdeterministicandnondeterministicpatterns.WealsotouchedupontheMarkovprocessandHiddenMarkovprocessingeneral.WecheckedtheclassesimplementedinsideMahouttosupporttheHiddenMarkovModel.WetookupanexampletocreatetheHMMmodelandfurtherusedthismodeltopredicttheobservationalstate’ssequence.WeusedtheViterbialgorithmimplementedinMahouttopredictthehiddenstatesinthesystem.

Now,inthenextchapter,wewillcoveronemoreinterestingalgorithmusedinclassificationarea:Randomforest.

Chapter6.LearningRandomForestUsingMahoutRandomforestisoneofthemostpopulartechniquesinclassification.Itstartswithamachinelearningtechniquecalleddecisiontree.Inthischapter,wewillexplorethefollowingtopics:

DecisiontreeRandomforestUsingMahoutforRandomforest

DecisiontreeAdecisiontreeisusedforclassificationandregressionproblems.Insimpleterms,itisapredictivemodelthatusesbinaryrulestocalculatethetargetvariable.Inadecisiontree,weuseaniterativeprocessofsplittingthedataintopartitions,thenwesplititfurtheronbranches.Asinotherclassificationmodelcreationprocesses,westartwiththetrainingdatasetinwhichtargetvariablesorclasslabelsaredefined.Thealgorithmtriestobreakalltherecordsintrainingdatasetsintotwopartsbasedononeoftheexplanatoryvariables.Thepartitioningisthenappliedtoeachnewpartition,andthisprocessiscontinueduntilnomorepartitioningcanbedone.Thecoreofthealgorithmistofindouttherulethatdeterminestheinitialsplit.Therearealgorithmstocreatedecisiontrees,suchasIterativeDichotomiser3(ID3),ClassificationandRegressionTree(CART),Chi-squaredAutomaticInteractionDetector(CHAID),andsoon.AgoodexplanationforID3canbefoundathttp://www.cse.unsw.edu.au/~billw/cs9414/notes/ml/06prop/id3/id3.html.

Formingtheexplanatoryvariablestochoosethebestsplitterinanode,thealgorithmconsiderseachvariableinturn.Everypossiblesplitisconsideredandtried,andthebestsplitistheonethatproducesthelargestdecreaseindiversityoftheclassificationlabelwithineachpartition.Thisisrepeatedforallvariables,andthewinnerischosenasthebestsplitterforthatnode.Theprocessiscontinuedinthenextnodeuntilwereachanodewherewecanmakethedecision.

Wecreateadecisiontreefromatrainingdatasetsoitcansufferfromtheoverfittingproblem.Thisbehaviorcreatesaproblemwithrealdatasets.Toimprovethissituation,aprocesscalledpruningisused.Inthisprocess,weremovethebranchesandleavesofthetreetoimprovetheperformance.Algorithmsusedtobuildthetreeworkbestatthestartingorrootnodesincealltheinformationisavailablethere.Lateron,witheachsplit,dataislessandtowardstheendofthetree,aparticularnodecanshowpatternsthatarerelatedtothesetofdatawhichisusedtosplit.Thesepatternscreateproblemswhenweusethemtopredicttherealdataset.Pruningmethodsletthetreegrowandremovethesmallerbranchesthatfailtogeneralize.Nowtakeanexampletounderstandthedecisiontree.

Considerwehaveairisflowerdataset.Thisdatasetishugelypopularinthemachinelearningfield.ItwasintroducedbySirRonaldFisher.Itcontains50samplesfromeachofthreespeciesofirisflower(Irissetosa,Irisvirginica,andIrisversicolor).Thefourexplanatoryvariablesarethelengthandwidthofthesepalsandpetalsincentimeters,andthetargetvariableistheclasstowhichtheflowerbelongs.

http://www.cse.unsw.edu.au/~billw/cs9414/notes/ml/06prop/id3/id3.html

Asyoucanseeintheprecedingdiagram,allthegroupswereearlierconsideredasSentosaspeciesandthentheexplanatoryvariableandpetallengthwerefurtherusedtodividethegroups.Ateachstep,thecalculationformisclassifieditemswasalsodone,whichshowshowmanyitemswerewronglyclassified.Moreover,thepetalwidthvariablewastakenintoaccount.Usually,itemsatleafnodesarecorrectlyclassified.

RandomforestTheRandomforestalgorithmwasdevelopedbyLeoBreimanandAdeleCutler.Randomforestsgrowmanyclassificationtrees.Theyareanensemblelearningmethodforclassificationandregressionthatconstructsanumberofdecisiontreesattrainingtimeandalsooutputstheclassthatisthemodeoftheclassesoutputtedbyindividualtrees.

Singledecisiontreesshowthebias–variancetradeoff.Sotheyusuallyhavehighvarianceorhighbias.Thefollowingaretheparametersinthealgorithm:

Bias:ThisisanerrorcausedbyanerroneousassumptioninthelearningalgorithmVariance:Thisisanerrorthatrangesfromsensitivitytosmallfluctuationsinthetrainingset

Randomforestsattempttomitigatethisproblembyaveragingtofindanaturalbalancebetweentwoextremes.ARandomforestworksontheideaofbagging,whichistoaveragenoisyandunbiasedmodelstocreateamodelwithlowvariance.ARandomforestalgorithmworksasalargecollectionofdecorrelateddecisiontrees.TounderstandtheideaofaRandomforestalgorithm,let’sworkwithanexample.

Considerwehaveatrainingdatasetthathaslotsoffeatures(explanatoryvariables)andtargetvariablesorclasses:

Wecreateasamplesetfromthegivendataset:

Adifferentsetofrandomfeaturesweretakenintoaccounttocreatetherandomsub-dataset.Now,fromthesesub-datasets,differentdecisiontreeswillbecreated.Soactuallywehavecreatedaforestofthedifferentdecisiontrees.Usingthesedifferenttrees,wewillcreatearankingsystemforalltheclassifiers.Topredicttheclassofanewunknownitem,wewilluseallthedecisiontreesandseparatelyfindoutwhichclassthesetreesare

predicting.Seethefollowingdiagramforabetterunderstandingofthisconcept:

Differentdecisiontreestopredicttheclassofanunknownitem

Inthisparticularcase,wehavefourdifferentdecisiontrees.Wepredicttheclassofanunknowndatasetwitheachofthetrees.Aspertheprecedingfigure,thefirstdecisiontreeprovidesclass2asthepredictedclass,theseconddecisiontreepredictsclass5,thethirddecisiontreepredictsclass5,andthefourthdecisiontreepredictsclass3.Now,aRandomforestwillvoteforeachclass.Sowehaveonevoteeachforclass2andclass3andtwovotesforclass5.Therefore,ithasdecidedthatforthenewunknowndataset,thepredictedclassisclass5.Sotheclassthatgetsahighervoteisdecidedforthenewdataset.ARandomforesthasalotofbenefitsinclassificationandafewofthemarementionedinthefollowinglist:

CombinationoflearningmodelsincreasestheaccuracyoftheclassificationRunseffectivelyonlargedatasetsaswellThegeneratedforestcanbesavedandusedforotherdatasetsaswellCanhandlealargeamountofexplanatoryvariables

NowthatwehaveunderstoodtheRandomforesttheoretically,let’smoveontoMahoutandusetheRandomforestalgorithm,whichisavailableinApacheMahout.

UsingMahoutforRandomforestMahouthasimplementationfortheRandomforestalgorithm.Itisveryeasytounderstandanduse.Solet’sgetstarted.

Dataset

WewillusetheNSL-KDDdataset.Since1999,KDD‘99hasbeenthemostwidelyuseddatasetfortheevaluationofanomalydetectionmethods.ThisdatasetispreparedbyS.J.StolfoandisbuiltbasedonthedatacapturedintheDARPA‘98IDSevaluationprogram(R.P.Lippmann,D.J.Fried,I.Graf,J.W.Haines,K.R.Kendall,D.McClung,D.Weber,S.E.Webster,D.Wyschogrod,R.K.Cunningham,andM.A.Zissman,“Evaluatingintrusiondetectionsystems:The1998darpaoff-lineintrusiondetectionevaluation,”discex,vol.02,p.1012,2000).

DARPA‘98isabout4GBofcompressedraw(binary)tcpdumpdataof7weeksofnetworktraffic,whichcanbeprocessedintoabout5millionconnectionrecords,eachwithabout100bytes.Thetwoweeksoftestdatahavearound2millionconnectionrecords.TheKDDtrainingdatasetconsistsofapproximately4,900,000singleconnectionvectors,eachofwhichcontains41featuresandislabeledaseithernormaloranattack,withexactlyonespecificattacktype.

NSL-KDDisadatasetsuggestedtosolvesomeoftheinherentproblemsoftheKDD‘99dataset.Youcandownloadthisdatasetfromhttp://nsl.cs.unb.ca/NSL-KDD/.

WewilldownloadtheKDDTrain+_20Percent.ARFFandKDDTest+.ARFFdatasets.

NoteInKDDTrain+_20Percent.ARFFandKDDTest+.ARFF,removethefirst44lines(that

http://nsl.cs.unb.ca/NSL-KDD/

is,alllinesstartingwith@attribute).Ifthisisnotdone,wewillnotbeabletogenerateadescriptorfile.

StepstousetheRandomforestalgorithminMahoutThestepstoimplementtheRandomforestalgorithminApacheMahoutareasfollows:

1. Transferthetestandtrainingdatasetstohdfsusingthefollowingcommands:

hadoopfs-mkdir/user/hue/KDDTrain

hadoopfs-mkdir/user/hue/KDDTest

hadoopfs–put/tmp/KDDTrain+_20Percent.arff/user/hue/KDDTrain

hadoopfs–put/tmp/KDDTest+.arff/user/hue/KDDTest

2. Generatethedescriptorfile.BeforeyoubuildaRandomforestmodelbasedonthetrainingdatainKDDTrain+.arff,adescriptorfileisrequired.Thisisbecauseallinformationinthetrainingdatasetneedstobelabeled.Fromthelabeleddataset,thealgorithmcanunderstandwhichoneisnumericalandcategorical.Usethefollowingcommandtogeneratedescriptorfile:

hadoopjar$MAHOUT_HOME/core/target/mahout-core-xyz.job.jar

org.apache.mahout.classifier.df.tools.Describe

-p/user/hue/KDDTrain/KDDTrain+_20Percent.arff

-f/user/hue/KDDTrain/KDDTrain+.info

-dN3C2NC4NC8N2C19NL

Jar:Mahoutcorejar(xyzstandsforversion).IfyouhavedirectlyinstalledMahout,itcanbefoundunderthe/usr/lib/mahoutfolder.ThemainclassDescribeisusedhereandittakesthreeparameters:

Theppathforthedatatobedescribed.

Theflocationforthegenerateddescriptorfile.

distheinformationfortheattributeonthedata.N3C2NC4NC8N2C19NLdefinesthatthedatasetisstartingwithanumeric(N),followedbythreecategoricalattributes,andsoon.Inthelast,Ldefinesthelabel.


3. BuildtheRandomforestusingthefollowingcommand:

hadoopjar$MAHOUT_HOME/examples/target/mahout-examples-xyz-job.jar

org.apache.mahout.classifier.df.mapreduce.BuildForest

-Dmapred.max.split.size=1874231-d

/user/hue/KDDTrain/KDDTrain+_20Percent.arff

-ds/user/hue/KDDTrain/KDDTrain+.info

-sl5-p-t100–o/user/hue/nsl-forest

Jar:Mahoutexamplejar(xyzstandsforversion).IfyouhavedirectlyinstalledMahout,itcanbefoundunderthe/usr/lib/mahoutfolder.Themainclassbuild

forestisusedtobuildtheforestwithotherarguments,whichareshowninthefollowinglist:

Dmapred.max.split.sizeindicatestoHadoopthemaximumsizeofeachpartition.

dstandsforthedatapath.

dsstandsforthelocationofthedescriptorfile.

slisavariabletoselectrandomlyateachtreenode.Here,eachtreeisbuiltusingfiverandomlyselectedattributespernode.

pusespartialdataimplementation.

tstandsforthenumberoftreestogrow.Here,thecommandsbuild100treesusingpartialimplementation.

ostandsfortheoutputpaththatwillcontainthedecisionforest.

Intheend,theprocesswillshowthefollowingresult:

4. Usethismodeltoclassifythenewdataset:

hadoopjar$MAHOUT_HOME/examples/target/mahout-examples-xyz-job.jar

org.apache.mahout.classifier.df.mapreduce.TestForest

-i/user/hue/KDDTest/KDDTest+.arff

-ds/user/hue/KDDTrain/KDDTrain+.info-m/user/hue/nsl-forest-a–mr

-o/user/hue/predictions

Jar:Mahoutexamplejar(xyzstandsforversion).IfyouhavedirectlyinstalledMahout,itcanbefoundunderthe/usr/lib/mahoutfolder.Theclasstotesttheforesthasthefollowingparameters:

Iindicatesthepathforthetestdata

dsstandsforthelocationofthedescriptorfile

mstandsforthelocationofthegeneratedforestfromthepreviouscommand

ainformstoruntheanalyzertocomputetheconfusionmatrix

mrinformshadooptodistributetheclassification

ostandsforthelocationtostorethepredictionsin

Thejobprovidesthefollowingconfusionmatrix:

So,fromtheconfusionmatrix,itisclearthat9,396instanceswerecorrectlyclassifiedand315normalinstanceswereincorrectlyclassifiedasanomalies.Andtheaccuracypercentageis77.7635(correctlyclassifiedinstancesbythemodel/classifiedinstances).Theoutputfileinthepredictionfoldercontainsthelistwhere0and1.0definesthenormaldatasetand1definestheanomaly.

SummaryInthischapter,wediscussedtheRandomforestalgorithm.WestartedourdiscussionbyunderstandingthedecisiontreeandcontinuedwithanunderstandingoftheRandomforest.WetookuptheNSL-KDDdataset,whichisusedtobuildpredictivesystemsforcybersecurity.WeusedMahouttobuildtheRandomforesttree,anduseditwiththetestdatasetandgeneratedtheconfusionmatrixandotherstatisticsfortheoutput.

Inthenextchapter,wewilllookatthefinalclassificationalgorithmavailableinApacheMahout.Sostaytuned!

Chapter7.LearningMultilayerPerceptronUsingMahoutTounderstandaMultilayerPerceptron(MLP),wewillfirstexploreonemorepopularmachinelearningtechnique:neuralnetwork.Inthischapter,wewillexplorethefollowingtopics:

NeuralnetworkandneuronsMLPUsingMahoutforMLPimplementation

NeuralnetworkandneuronsNeuralnetworkisanoldalgorithm,anditwasdevelopedwithagoalinmind:toprovidethecomputerwithabrain.Neuralnetworkisinspiredbythebiologicalstructureofthehumanbrainwheremultipleneuronsareconnectedandformcolumnsandlayers.Aneuronisanelectricallyexcitablecellthatprocessesandtransmitsinformationthroughelectricalandchemicalsignals.Perceptualinputentersintotheneuralnetworkthroughoursensoryorgansandisthenfurtherprocessedintohigherlevels.Let’sunderstandhowneuronsworkinourbrain.

Neuronsarecomputationalunitsinthebrainthatcollecttheinputfrominputnerves,whicharecalleddendrites.Theyperformcomputationontheseinputmessagesandsendtheoutputusingoutputnerves,whicharecalledaxons.Seethefollowingfigure(http://vv.carleton.ca/~neil/neural/neuron-a.html):

Onthesamelines,wedevelopaneuralnetworkincomputers.Wecanrepresentaneuroninouralgorithmasshowninthefollowingfigure:

http://vv.carleton.ca/~neil/neural/neuron-a.html

Here,x1,x2,andx3arethefeaturevectors,andtheyareassignedtoafunctionf,whichwilldothecomputationandprovidetheoutput.Thisactivationfunctionisusuallychosenfromthefamilyofsigmoidalfunctions(asdefinedinChapter3,LearningLogisticRegression/SGDUsingMahout).Inthecaseofclassificationproblems,softmaxactivationfunctionsareused.Inclassificationproblems,wewanttheoutputastheprobabilitiesoftargetclasses.So,itisdesirablefortheoutputtoliebetween0and1andthesumcloseto1.Softmaxfunctionenforcestheseconstraints.Itisageneralizationofthelogisticfunction.Moredetailsonsoftmaxfunctioncanbefoundathttp://www.faqs.org/faqs/ai-faq/neural-nets/part2/section-12.html.

http://www.faqs.org/faqs/ai-faq/neural-nets/part2/section-12.html

MultilayerPerceptronAneuralnetworkorartificialneuralnetworkgenerallyreferstoanMLPnetwork.Wedefinedneuronasanimplementationincomputersintheprevioussection.AnMLPnetworkconsistsofmultiplelayersoftheseneuronunits.Let’sunderstandaperceptronnetworkofthreelayers,asshowninthenextfigure.ThefirstlayeroftheMLPrepresentstheinputandhasnootherpurposethanroutingtheinputtoeveryconnectedunitinafeed-forwardfashion.Thesecondlayeriscalledhiddenlayers,andthelastlayerservesthespecialpurposeofdeterminingtheoutput.Theactivationofneuronsinthehiddenlayerscanbedefinedasthesumoftheweightofalltheinput.Neuron1inlayer2isdefinedasfollows:

Y12=g(w110x0+w111x1+w112x2+w113x3)

Thefirstpartwhere*x0=0*iscalledthebiasandcanbeusedasanoffset,independentoftheinput.Neuron2inlayer2isdefinedasfollows:

Y22=g(w120x0+w121x1+w122x2+w123x3)

Neuron3inlayer2isdefinedasfollows:

Y32=g(w130x0+w131x1+w132x2+w133x3)

Here,gisasigmoidfunction,asdefinedinChapter3,LearningLogisticRegression/SGDUsingMahout.Thefunctionisasfollows:

g(z)=1/1+e(-z)

InthisMLPnetworkoutput,fromeachinputandhiddenlayers,neuronunitsaredistributedtoothernodes,andthisiswhythistypeofnetworkiscalledafullyconnectednetwork.Inthisnetwork,novaluesarefedbacktothepreviouslayer.(Feedforwardis

anotherstrategyandisalsoknownasbackpropagation.Detailsonthiscanbefoundathttp://home.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html.)

AnMLPnetworkcanhavemorethanonehiddenlayer.TogetthevalueoftheweightssothatwecangetthepredictedvalueascloseaspossibletotheactualoneisatrainingprocessoftheMLP.Tobuildaneffectivenetwork,weconsideralotofitemssuchasthenumberofhiddenlayersandneuronunitsineachlayer,thecostfunctiontominimizetheerrorinpredictedandactualvalues,andsoon.

Nowlet’sdiscusstwomoreimportantandproblematicquestionsthatarisewhencreatinganMLPnetwork:

Howmanyhiddenlayersshouldoneuseforthenetwork?Howmanynumbersofhiddenunits(neuronunits)shouldoneuseinahiddenlayer?

Zerohiddenlayersarerequiredtoresolvelinearlyseparabledata.Assumingyourdatadoesrequireseparationbyanon-lineartechnique,alwaysstartwithonehiddenlayer.Almostcertainly,that’sallyouwillneed.IfyourdataisseparableusinganMLP,thenthisMLPprobablyonlyneedsasinglehiddenlayer.Inordertoselectthenumberofunitsindifferentlayers,thesearetheguidelines:

Inputlayer:Thisreferstothenumberofexplanatoryvariablesinthemodelplusoneforthebiasnode.Outputlayer:Inthecaseofclassification,thisreferstothenumberoftargetvariables,andinthecaseofregression,thisisobviouslyone.Hiddenlayer:Startyournetworkwithonehiddenlayerandusethenumberofneuronunitsequivalenttotheunitsintheinputlayer.Thebestwayistotrainseveralneuralnetworkswithdifferentnumbersofhiddenlayersandhiddenneuronsandmeasuretheperformanceofthesenetworksusingcross-validation.Youcanstickwiththenumberthatyieldsthebest-performingnetwork.Problemsthatrequiretwohiddenlayersarerarelyencountered.However,neuralnetworksthathavemorethanonehiddenlayercanrepresentfunctionswithanykindofshape.Thereiscurrentlynotheorytojustifytheuseofneuralnetworkswithmorethantwohiddenlayers.Infact,formanypracticalproblems,thereisnoreasontouseanymorethanonehiddenlayer.Anetworkwithnohiddenlayerisonlycapableofrepresentinglinearlyseparablefunctions.Networkswithonelayercanapproximateanyfunctionthatcontainsacontinuousmappingfromonefinitespacetoanother,andnetworkswithtwohiddenlayerscanrepresentanarbitrarydecisionboundarytoarbitraryaccuracywithrationalactivationfunctionsandcanapproximateanysmoothmappingtoanyaccuracy(Chapter5ofthebookIntroductiontoNeuralNetworksforJava).Numberofneuronsorhiddenunits:Usethenumberofneuronunitsequivalenttotheunitsintheinputlayer.Thenumberofhiddenunitsshouldbelessthantwicethenumberofunitsintheinputlayer.Anotherruletocalculatethisis(numberofinputunits+numberofoutputunits)*2/3.

Dothetestingforgeneralizationerrors,trainingerrors,bias,andvariance.Whenageneralizationerrordips,thenjustbeforeitbeginstoincreaseagain,thenumbersofnodesareusuallyfoundtobeperfectatthispoint.

http://home.agh.edu.pl/~vlsi/AI/backp_t_en/backprop.html

Nowlet’smoveontothenextsectionandexplorehowwecanuseMahoutforanMLP.

MLPimplementationinMahoutTheMLPimplementationisbasedonamoregeneralneuralnetworkclass.ItisimplementedtorunonasinglemachineusingStochasticGradientDescent,wheretheweightsareupdatedusingonedatapointatatime.

Thenumberoflayersandunitsperlayercanbespecifiedmanuallyanddeterminesthewholetopologywitheachunitbeingfullyconnectedtothepreviouslayer.Abiasunitisautomaticallyaddedtotheinputofeverylayer.Abiasunitishelpfulforshiftingtheactivationfunctiontotheleftorright.Itislikeaddingacoefficienttothelinearfunction.

Currently,thelogisticsigmoidisusedasasquashingfunctionineveryhiddenandoutputlayer.

Thecommand-lineversiondoesnotperformiterationsthatleadtobadresultsonsmalldatasets.AnotherrestrictionisthattheCLIversionoftheMLPonlysupportsclassification,sincethelabelshavetobegivenexplicitlywhenexecutingtheimplementationinthecommandline.

Alearnedmodelcanbestoredandupdatedwithnewtraininginstancesusingthe`--update`flag.Theoutputoftheclassificationresultissavedasa.txtfileandonlyconsistsoftheassignedlabels.Apartfromthecommand-lineinterface,itispossibletoconstructandcompilemorespecializedneuralnetworksusingtheAPIandinterfacesinthemrlegacypackage.(Thecorepackageisrenamedasmrlegacy.)

Inthecommandline,weuseTrainMultilayerPerceptronandRunMultilayerPerceptronclassesthatareavailableinthemrlegacypackagewiththreeotherclasses:Neuralnetwork.java,NeuralNetworkFunctions.java,andMultilayerPerceptron.java.Forthisparticularimplementation,userscanfreelycontrolthetopologyoftheMLP,includingthefollowing:

ThesizeoftheinputlayerThenumberofhiddenlayersThesizeofeachhiddenlayerThesizeoftheoutputlayerThecostfunctionThesquashingfunction

Themodelistrainedinanonlinelearningapproach,wheretheweightsofneuronsintheMLPisupdatedandincrementedusingthebackPropagationalgorithmproposedbyRumelhart,D.E.,Hinton,G.E.,andWilliams,R.J.(1986),Learningrepresentationsbyback-propagatingerrors.Nature,323,533-536.

UsingMahoutforMLPMahouthasimplementationforanMLPnetwork.TheMLPimplementationiscurrentlylocatedintheMap-Reduce-Legacypackage.Aswithotherclassificationalgorithms,twoseparatedclassesareimplementedtotrainandusethisclassifier.Fortrainingtheclassifier,theorg.apache.mahout.classifier.mlp.TrainMultilayerPerceptronclass,andforrunningtheclassifier,theorg.apache.mahout.classifier.mlp.RunMultilayerPerceptronclassisused.Thereareanumberofparametersdefinedthatareusedwiththeseclasses,butwewilldiscusstheseparametersoncewerunourexampleonadataset.

Dataset

Inthischapter,wewilltrainanMLPtoclassifytheirisdataset.Theirisflowerdatasetcontainsdataofthreeflowerspecies,whereeachdatapointconsistsoffourfeatures.ThisdatasetwasintroducedbySirRonaldFisher.Itconsistsof50samplesfromeachofthreespeciesofiris.ThesespeciesareIrissetosa,Irisvirginica,andIrisversicolor.Fourfeaturesweremeasuredfromeachsample:

SepallengthSepalwidthPetallengthPetalwidth

Allmeasurementsareincentimeters.Youcandownloadthisdatasetfromhttps://archive.ics.uci.edu/ml/machine-learning-databases/iris/andsaveitasa.csvfile,asshowninthefollowingscreenshot:

Thisdatasetwilllooklikethethefollowingscreenshot:

https://archive.ics.uci.edu/ml/machine-learning-databases/iris/

StepstousetheMLPalgorithminMahoutThestepstousetheMLPalgorithminMahoutareasfollows:

1. CreatetheMLPmodel.

TocreatetheMLPmodel,wewillusetheTrainMultilayerPerceptronclass.Usethefollowingcommandtogeneratethemodel:

bin/mahoutorg.apache.mahout.classifier.mlp.TrainMultilayerPerceptron-

i/tmp/irisdata.csv-labelsIris-setosaIris-versicolorIris-virginica

-mo/tmp/model.model-ls483-l0.2-m0.35-r0.0001

Youcanalsorunusingthecorejar:Mahoutcorejar(xyzstandsfortheversion).IfyouhavedirectlyinstalledMahout,itcanbefoundunderthe/usr/lib/mahoutfolder.Executethefollowingcommand:

Java–cp/usr/lib/mahout/mahout-core-xyz-job.jar

org.apache.mahout.classifier.mlp.TrainMultilayerPerceptron-i

/tmp/irisdata.csv-labelsIris-setosaIris-versicolorIris-virginica-

mo/user/hue/mlp/model.model-ls483-l0.2-m0.35-r0.0001

TheTrainMultilayerPerceptronclassisusedhereandittakesdifferentparameters.Also,iisthepathfortheinputdataset.Here,wehaveputthedatasetunderthe/tmpfolder(localfilesystem).Additionally,labelsaredefinedinthedataset.Herewehavethefollowinglabels:

moistheoutputlocationforthecreatedmodel.lsisthenumberofunitsperlayer,includinginput,hidden,andoutputlayers.Thisparameterspecifiesthetopologyofthenetwork.Here,wehave4astheinputfeature,8forthehiddenlayer,and3fortheoutputclassnumber.listhelearningratethatisusedforweightupdates.Thedefaultis0.5.Toapproximategradientdescent,neuralnetworksaretrainedwithalgorithms.Learningispossibleeitherbybatchoronlinemethods.Inbatchtraining,weightchangesareaccumulatedoveranentirepresentationofthetrainingdata(anepoch)beforebeingapplied,whileonlinetrainingupdatesweighsafterthepresentationofeachtrainingexample(instance).Moredetailscanbefoundathttp://axon.cs.byu.edu/papers/Wilson.nn03.batch.pdf.misthemomentumweightthatisusedforgradientdescent.Thismustbeintherangebetween0–1.0.ristheregularizationvaluefortheweightvector.Thismustbeintherangebetween0–0.1.Itisusedtopreventoverfitting.

http://axon.cs.byu.edu/papers/Wilson.nn03.batch.pdf

2. Totest/runtheMLPclassificationofthetrainedmodel,wecanusethefollowingcommand:

bin/mahoutorg.apache.mahout.classifier.mlp.RunMultilayerPerceptron-i

/tmp/irisdata.csv-cr03-mo/tmp/model.model-o/tmp/labelResult.txt

YoucanalsorunusingtheMahoutcorejar(xyzstandsforversion).IfyouhavedirectlyinstalledMahout,itcanbefoundunderthe/usr/lib/mahoutfolder.Executethefollowingcommand:

Java–cp/usr/lib/mahout/mahout-core-xyz-job.jar

org.apache.mahout.classifier.mlp.RunMultilayerPerceptron-i

/tmp/irisdata.csv-cr03-mo/tmp/model.model-o/tmp/labelResult.txt

TheRunMultilayerPerceptronclassisemployedheretousethemodel.Thisclassalsotakesdifferentparameters,whichareasfollows:

iindicatestheinputdatasetlocationcristherangeofcolumnstousefromtheinputfile,startingwith0(thatis,`-cr05`forincludingthefirstsixcolumnsonly)moisthelocationofthemodelbuiltearlieroisthepathtostorelabeledresultsfromrunningthemodel

SummaryInthischapter,wediscussedoneofthenewlyimplementedalgorithmsinMahout:MLP.WestartedourdiscussionbyunderstandingneuralnetworksandneuronunitsandcontinuedourdiscussionfurthertounderstandtheMLPnetworkalgorithm.Wediscussedhowtochoosedifferentlayerunits.WethenmovedtoMahoutandusedtheirisdatasettotestandrunanMLPalgorithmimplementedinMahout.Withthis,wehavefinishedourdiscussiononclassificationalgorithmsavailableinApacheMahout.

NowwemoveontothenextchapterofthisbookwherewewilldiscussthenewchangescomingupinthenewMahoutrelease.

Chapter8.MahoutChangesintheUpcomingReleaseMahoutisacommunity-drivenprojectanditscommunityisverystrong.Thiscommunitydecidedonsomeofthemajorchangesintheupcoming1.0release.Inthischapter,wewillexploretheupcomingchangesanddevelopmentsinApacheMahout.Wewilllookatthefollowingtopicsinbrief:

NewchangesdueinMahout1.0ApacheSparkH20-platform-relatedworkinApacheMahout

MahoutnewchangesMahoutwasusingthemapreduceprogrammingmodeltohandlelargedatasets.FromtheendofApril2014,thecommunitydecidedtostoptheimplementationofthenewmapreducealgorithm.Thisdecisionhasavalidreason.Mahout’scodebasewillbemovingtomoderndataprocessingsystemsthatofferaricherprogrammingmodelandmoreefficientexecutionthanHadoop’sMapReduce.

MahouthasstarteditsimplementationonthetopofDomainSpecificLanguage(DSL)forlinearalgebraicoperations.ProgramswritteninthisDSLareautomaticallyoptimizedandexecutedinparallelonApacheSpark.ScalaDSLandalgebraicoptimizerisScalaandSparkbindingforMahout.

MahoutScalaandSparkbindingsWithMahoutScalabindingsandMahoutSparkbindingsforlinearalgebrasubroutines,developersinMahoutaretryingtobringsemanticexplicitnesstoMahout’sin-coreandout-of-corelinearalgebrasubroutines.TheyaredoingthiswhileaddingthebenefitsofthestrongprogrammingenvironmentofScalaandcapitalizingonscalabilitybenefitsofSparkandGraphX.ScalabindingisusedtoprovidesupportforScalaDSL,andthiswillmakewritingmachinelearningprogramseasier.

MahoutScalaandSparkbindingsarepackagesthataimtoprovideanR-likelookandfeeltoMahout’sin-coreandout-of-coreSpark-backedlinearalgebra.AnimportantpartofSparkbindingsistheexpressionoptimizer.Thisoptimizerlooksattheentireexpressionanddecidesonhowitcanbesimplifiedandwhichphysicaloperatorsshouldbepicked.Ahigh-leveldiagramofthebindingstackisshowninthefollowingfigure(https://issues.apache.org/jira/secure/attachment/12638098/BindingsStack.jpg):

TheSparkbindingshellhasalsobeenimplementedinMahout1.0.Let’sunderstandtheApacheSparkprojectfirstandthenwewillrevisittheSparkbindingshellinMahout.

https://issues.apache.org/jira/secure/attachment/12638098/BindingsStack.jpg

ApacheSparkApacheSparkisanopensource,in-memory,general-purposecomputingsystem.Spark’sin-memorytechniqueprovidesperformancethatis100timesfaster.InsteadofHadoop-likedisk-basedcomputation,Sparkusesclustermemorytouploadallthedataintothememory,andthisdatacanbequeriedrepeatedly.

ApacheSparkprovideshigh-levelAPIsinJava,Python,andScalaandanoptimizedenginethatsupportsgeneralexecutiongraphs.Itprovidesthefollowinghigh-leveltools:

SparkSQL:ThisisforSQLandstructureddataprocessing.MLib:ThisisSpark’sscalablemachinelearninglibrarythatconsistsofcommonlearningalgorithmsandutilities,includingclassification,regression,clustering,collaborativefiltering,dimensionalityreduction,aswellastheunderlyingoptimizationprimitives.GraphX:ThisisthenewSparkAPIforgraphsandgraph-parallelcomputation.Sparkstreaming:Thiscancollectdatafrommanysourcesandafterprocessingthisdata,itusescomplexalgorithmsandcanpushthedatatofilesystems,databases,andlivedashboards.

AsSparkisgainingpopularityamongdatascientists,theMahoutcommunityisalsoquicklyworkingonmakingMahoutalgorithmsfunctiononSpark’sexecutionenginetospeedupitscalculation10to100timesfaster.MahoutprovidesseveralimportantbuildingblockstocreaterecommendationsusingSpark.Spark-itemsimilaritycanbeusedtocreateotherpeoplealsolikedthesethingskindofrecommendationsandwhenpairedwithasearchenginecanpersonalizerecommendationsforindividualusers.Spark-rowsimilaritycanprovidenon-personalizedcontentbasedonrecommendationsandwhenpairedwithasearchenginecanbeusedtopersonalizecontentbasedonrecommendations(http://comments.gmane.org/gmane.comp.apache.mahout.scm/6513).

http://comments.gmane.org/gmane.comp.apache.mahout.scm/6513

UsingMahout’sSparkshellYoucanuseMahout’sSparkshellbyreferringtothefollowingsteps:

1. DownloadSparkfromhttp://spark.apache.org/downloads.html.2. Createanewfolderwiththenamesparkusingthefollowingcommandandmovethe

downloadedfilethere:

mkdir/tmp/spark

mv~/Downloads/spark-1.1.1.tgz/tmp/spark

3. Unpackthearchivedfileinafolderusingthefollowingcommand:

cd/tmp/spark

tarxzfspark-1.1.1.tgz

4. Thiswillunzipthefileunder/tmp/spark/spark-1.1.1.Now,movetothenewlycreatedfolderandrunthefollowingcommand:

cd/spark-1.1.1

sbt/sbtassembly

ThiswillbuildSparkonyoursystemasshowninthefollowingscreenshot:

5. NowcreateaMahoutdirectoryandmovethefiletoitusingthefollowingcommand:

mkdir/tmp/Mahout

6. CheckoutthemasterbranchofMahoutfromGitHubusingthefollowingcommand:

gitclonehttps://github.com/apache/mahoutmahout

http://spark.apache.org/downloads.html

Theoutputoftheprecedingcommandisshowninthefollowingscreenshot:

7. ChangeyourdirectorytothenewlycreatedMahoutdirectoryandbuildMahout:

cdmahout

mvn-DskipTestscleaninstall


8. MovetothedirectorywhereyouunpackedSparkandtypethefollowingcommandtostartSparklocally:

cd/tmp/spark/spark-1.1.1

sbin/start-all-sh


9. Openabrowser;pointittohttp://localhost:8080/tocheckwhetherSparkhas

successfullystarted.CopytheURLoftheSparkmasteratthetopofthepage(itstartswithspark://).

10. Definethefollowingenvironmentvariables:

exportMAHOUT_HOME=[directoryintowhichyoucheckedoutMahout]

exportSPARK_HOME=[directorywhereyouunpackedSpark]

exportMASTER=[urloftheSparkmaster]

11. Finally,changetothedirectorywhereyouunpackedMahoutandtypebin/mahoutspark-shell;youshouldseetheshellstartingandgetthemahout>prompt.

NowyourMahoutSparkshellisreadyandyoucanstartplayingwithdata.Formoreinformationonthistopic,seetheimplementationsectionathttps://mahout.apache.org/users/sparkbindings/play-with-shell.html.

https://mahout.apache.org/users/sparkbindings/play-with-shell.html

H2OplatformintegrationAsdiscussedearlier,anexperimentalworktointegrateMahoutandtheH2Oplatformisalsoinprogress.TheintegrationprovidesanH2ObackendtotheMahoutalgebraDSL.

H2OmakesHadoopdomath!H2Oscalesstatistics,machinelearning,andmathoverbigdata.Itisextensibleanduserscanbuildblocksusingsimplemathlegosinthecore.H2OkeepsfamiliarinterfacessuchasR,Excel,andJSONsothatbigdataenthusiastsandexpertscanexplore,munge,model,andscoredatasetsusingarangeofsimple-to-advancedalgorithms.Datacollectioniseasy,whiledecisionmakingishard.H2Omakesitfastandeasytoderiveinsightsfromyourdatathroughfasterandbetterpredictivemodeling.Italsohasavisionofonlinescoringandmodelinginasingleplatform(http://0xdata.com/download/).

H2Oisfundamentallyapeer-to-peersystem.H2Onodesjointogethertoformacloudonwhichhigh-performancedistributedmathcanbeexecuted.Eachnodejoinsacloudofagivenname.Multiplecloudscanexistonthesamenetworkatthesametimeaslongastheirnamesaredifferent.Multiplenodescanexistonthesameserveraswell(theycanevenbelongtothesamecloud).

TheMahoutH2OintegrationisfitintothismodelbyhavingN-1workernodesandonedrivernode,allbelongingtothesamecloudname.Thedefaultcloudnameusedfortheintegrationismah2out.Cloudshavetobespunupaspertheirtask/job.

Moredetailscanbefoundathttps://issues.apache.org/jira/browse/MAHOUT-1500.

http://0xdata.com/download/

https://issues.apache.org/jira/browse/MAHOUT-1500

SummaryInthischapter,wediscussedtheupcomingreleaseofMahout1.0,andthechangesthatarecurrentlygoingon.WealsoglancedthroughSpark,Scalabinding,andApacheSpark.Wealsodiscussedahigh-leveloverviewofH2OMahoutintegration.

Nowlet’smoveontothefinalchapterofthisbookwherewewilldevelopaproduction-readyclassifier.

Chapter9.BuildinganE-mailClassificationSystemUsingApacheMahoutInthischapter,wewillcreateaclassifiersystemusingMahout.Inordertobuildthissystem,wewillcoverthefollowingtopics:

GettingthedatasetPreparationofthedatasetPreparingthemodelTrainingthemodel

Inthischapter,wewilltargetthecreationoftwodifferentclassifiers.Thefirstonewillbeaneasyonebecauseyoucanbothcreateandtestitonapseudo-distributedHadoopinstallation.Forthesecondclassifier,Iwillprovideyouwithallthedetails,soyoucanrunitusingyourfullydistributedHadoopinstallation.Iwillcountthesecondoneasahands-onexerciseforthereadersofthisbook.

Firstofall,let’sunderstandtheproblemstatementforthefirstusecase.Nowadays,inmostofthee-mailsystems,weseethate-mailsareclassifiedasspamornotspam.E-mailsthatarenotspamaredelivereddirectlyintoourinboxbutspame-mailsarestoredinafoldercalledSpam.Usually,basedonacertainpatternsuchasmessagesubject,sender’se-mailaddress,orcertainkeywordsinthemessagebody,wecategorizeanincominge-mailasspam.WewillcreateaclassifierusingMahout,whichwillclassifyane-mailintospamornotspam.WewilluseSpamAssassin,anApacheopensourceprojectdatasetforthistask.

Forthesecondusecase,wewillcreateaclassifier,whichcanpredictagroupofincominge-mails.Asanopensourceproject,therearelotsofprojectsundertheApachesoftwarefoundation,suchasApacheMahout,ApacheHadoop,ApacheSolr,andsoon.WewilltaketheApacheSoftwareFoundation(ASF)e-maildatasetandusingthis,wewillcreateandtrainourmodelsothatourmodelcanpredictanewincominge-mail.So,basedoncertainfeatures,wewillbeabletopredictwhichgroupanewincominge-mailbelongsto.

InMahout’sclassificationproblem,wewillhavetoidentifyapatterninthedatasettohelpuspredictthegroupofanewe-mail.Wealreadyhaveadataset,whichisseparatedbyprojectnames.WewillusetheASFpublice-mailarchivesdatasetforthisusecase.

Now,let’sconsiderourfirstusecase:spame-maildetectionclassifier.

Spame-maildatasetAsImentioned,wewillbeusingtheApacheSpamAssassinprojectsdataset.ApacheSpamAssassinisanopensourcespamfilter.Download20021010_easy_ham.tarand20021010_spam.tarfromhttp://spamassassin.apache.org/publiccorpus/,asshowninthefollowingscreenshot:

http://spamassassin.apache.org/publiccorpus/

CreatingthemodelusingtheAssassindatasetWecancreatethemodelwiththehelpofthefollowingsteps:

1. Createafolderundertmpwiththenamedataset,andthenclickonthefolderandunzipthedatasetsusingthefollowingcommand:

mkdir/tmp/assassin/dataset

tar–xvf/tmp/assassin/20021010_easy_ham.tar.bz2

tar–xvf/tmp/assassin/20021010_spam.tar.bz2

Thiswillcreatetwofoldersunderthedatasetfolder,easy_hamandspam,asshowninthefollowingscreenshot:

2. CreateafolderinHdfsandmovethisdatasetintoHadoop:

hadoopfs-mkdir/user/hue/assassin/

hadoopfs–put/tmp/assassin/dataset/user/hue/assassin

tar–xvf/tmp/assassin/20021010_spam.tar.bz2

Nowourdatapreparationisdone.Wehavedownloadedthedataandmovedthisdataintohdfs.Let’smoveontothenextstep.

3. ConvertthisdataintosequencefilessothatwecanprocessitusingHadoop:

bin/mahoutseqdirectory–i/user/hue/assassin/dataset–o

/user/hue/assassinseq-out

4. Convertthesequencefileintosparsevector(Mahoutalgorithmsacceptinputinvectorformat,whichiswhyweareconvertingthesequencefileintosparsevector)byusingthefollowingcommand:

bin/mahoutseq2sparse-i/user/hue/assassinseq-out/part-m-00000-o

/user/hue/assassinvec-lnorm-nv-wttfidf

Thecommandintheprecedingscreenshotisexplainedasfollows:

lnorm:Thiscommandisusedforoutputvectortobelognormalized.nv:Thiscommandisusedfornamedvector.wt:Thiscommandisusedtoidentifythekindofweighttouse.Hereweusetf-idf.

5. Splitthesetofvectorsfortrainingandtestingthemodel,asfollows:

bin/mahoutsplit-i/user/hue/assassinvec/tfidf-vectors--

trainingOutput/user/hue/assassindatatrain--testOutput

/user/hue/assassindatatest--randomSelectionPct20--overwrite--

sequenceFiles-xmsequential

Theprecedingcommandcanbeexplainedasfollows:

TherandomSelectionPctparameterdividesthepercentageofdataintotestandtrainingdatasets.Inthiscase,it’s80percentfortestand20percentfortraining.Thexmparameterspecifieswhatportionofthetf(tf-idf)vectorsistobeusedexpressedintimesthestandarddeviation.Thesigmasymbolspecifiesthedocumentfrequenciesofthesevectors.Itcanbeusedtoremovereallyhighfrequencyterms.Itisexpressedasadoublevalue.Agoodvaluetobespecifiedis3.0.Ifthevalueislessthan0,novectorswillbefilteredout.

6. Now,trainthemodelusingthefollowingcommand:

bin/mahouttrainnb-i/user/hue/assassindatatrain-el-o

/user/hue/prodmodel-li/user/hue/prodlabelindex-ow-c

7. Now,testthemodelusingthefollowingcommand:

bin/mahouttestnb-i/user/hue/assassindatatest-m/user/hue/prodmodel/

-l/user/hue/prodlabelindex-ow-o/user/hue/prodresults

Youcanseefromtheresultsthattheoutputisdisplayedontheconsole.Asperthematrix,thesystemhascorrectlyclassified99.53percentoftheinstancesgiven.

Wecanusethiscreatedmodeltoclassifynewdocuments.Todothis,wecaneitheruseaJavaprogramorcreateaservletthatcanbedeployedonourserver.

Let’stakeanexampleofaJavaprogramincontinuationofthisexercise.

ProgramtouseaclassifiermodelWewillcreateaJavaprogramthatwilluseourmodeltoclassifynewe-mails.Thisprogramwilltakemodel,labelindex,dictionary-file,documentfrequency,andtextfileasinputandwillgenerateascoreforthecategories.Thecategorywillbedecidedbasedonthehigherscores.

Let’shavealookatthisprogramstepbystep:

The.jarfilesrequiredtomakeacompilationofthisprogramareasfollows:

Hadoop-core-x.y.x.jar

Mahout-core-xyz.jar

Mahout-integration-xyz.jar

Mahout-math-xyz.jar

Theimportstatementsarelistedasfollows.WearediscussingthisbecausetherearelotsofchangesintheMahoutreleasesandpeopleusuallyfinditdifficulttogetthecorrectclasses.

importjava.io.BufferedReader;

importjava.io.FileReader;

importjava.io.StringReader;

importjava.util.HashMap;

importjava.util.Map;

importorg.apache.hadoop.conf.Configuration;

importorg.apache.hadoop.fs.Path;

importorg.apache.lucene.analysis.Analyzer;

importorg.apache.lucene.analysis.TokenStream;

importorg.apache.lucene.analysis.standard.StandardAnalyzer;

import

org.apache.lucene.analysis.tokenattributes.CharTermAttribute;

importorg.apache.lucene.util.Version;

importorg.apache.mahout.classifier.naivebayes.BayesUtils;

importorg.apache.mahout.classifier.naivebayes.NaiveBayesModel;

import

org.apache.mahout.classifier.naivebayes.StandardNaiveBayesClassifier;

importorg.apache.mahout.common.Pair;

import

org.apache.mahout.common.iterator.sequencefile.SequenceFileIterable;

importorg.apache.mahout.math.RandomAccessSparseVector;

importorg.apache.mahout.math.Vector;

importorg.apache.mahout.math.Vector.Element;

importorg.apache.mahout.vectorizer.TFIDF;

importorg.apache.hadoop.io.*;

importcom.google.common.collect.ConcurrentHashMultiset;

importcom.google.common.collect.Multiset;

Thesupportingmethodstoreadthedictionaryareasfollows:






dictionary.put(pair.getFirst().toString(),pair.getSecond().get());

}

returndictionary;

}

Thesupportingmethodstoreadthedocumentfrequencyareasfollows:

publicstaticMap<Integer,Long>readDocumentFrequency(Configuration

conf,PathdocumentFrequencyPath){

Map<Integer,Long>documentFrequency=newHashMap<Integer,Long>();

for(Pair<IntWritable,LongWritable>pair:new

SequenceFileIterable<IntWritable,LongWritable>(documentFrequencyPath,

true,conf)){

documentFrequency.put(pair.getFirst().get(),


}

returndocumentFrequency;

}

Thefirstpartofthemainmethodisusedtoperformthefollowingactions:

GettingtheinputLoadingthemodelInitializingStandardNaiveBayesClassifierusingourcreatedmodelReadinglabelindex,documentfrequency,anddictionarycreatedwhilecreatingthevectorfromthedataset

Thefollowingcodecanbeusedfortheprecedingactions:

publicstaticvoidmain(String[]args)throwsException{

if(args.length<5){

System.out.println("Arguments:[model][labelindex]

[dictionary][documentfrequency][newfile]");

return;

}

StringmodelPath=args[0];

StringlabelIndexPath=args[1];

StringdictionaryPath=args[2];

StringdocumentFrequencyPath=args[3];

StringnewDataPath=args[4];

Configurationconfiguration=newConfiguration();//modelisa

matrix(wordId,labelId)=>probabilityscore

NaiveBayesModelmodel=NaiveBayesModel.materialize(new

Path(modelPath),configuration);

StandardNaiveBayesClassifierclassifier=new

StandardNaiveBayesClassifier(model);

//labelsisamaplabel=>classId

Map<Integer,String>labels=

BayesUtils.readLabelIndex(configuration,newPath(labelIndexPath));

Map<String,Integer>dictionary=readDictionary(configuration,

newPath(dictionaryPath));

Map<Integer,Long>documentFrequency=

readDocumentFrequency(configuration,new

Path(documentFrequencyPath));

Thesecondpartofthemainmethodisusedtoextractwordsfromthee-mail:

Analyzeranalyzer=newStandardAnalyzer(Version.LUCENE_CURRENT);

intlabelCount=labels.size();

intdocumentCount=documentFrequency.get(-1).intValue();

System.out.println("Numberoflabels:"+labelCount);

System.out.println("Numberofdocumentsintrainingset:"+

documentCount);

BufferedReaderreader=newBufferedReader(new

FileReader(newDataPath));

while(true){

Stringline=reader.readLine();

if(line==null){

break;

}

ConcurrentHashMultiset<Object>words=

ConcurrentHashMultiset.create();

//extractwordsfrommail

TokenStreamts=analyzer.tokenStream("text",new

StringReader(line));

CharTermAttributetermAtt=ts.addAttribute(CharTermAttribute.class);

ts.reset();

intwordCount=0;

while(ts.incrementToken()){

if(termAtt.length()>0){

Stringword=

ts.getAttribute(CharTermAttribute.class).toString();

IntegerwordId=dictionary.get(word);

//ifthewordisnotinthedictionary,skipit

if(wordId!=null){

words.add(word);

wordCount++;

}

}

}

ts.close();

Thethirdpartofthemainmethodisusedtocreatevectoroftheidwordandthetf-idfweights:

Vectorvector=newRandomAccessSparseVector(10000);

TFIDFtfidf=newTFIDF();

for(Multiset.Entryentry:words.entrySet()){

Stringword=(String)entry.getElement();

intcount=entry.getCount();

IntegerwordId=dictionary.get(word);

Longfreq=documentFrequency.get(wordId);

doubletfIdfValue=tfidf.calculate(count,freq.intValue(),

wordCount,documentCount);

vector.setQuick(wordId,tfIdfValue);

}

Inthefourthpartofthemainmethod,withclassifier,wegetthescoreforeachlabelandassignthee-mailtothehigherscoredlabel:

VectorresultVector=classifier.classifyFull(vector);

doublebestScore=-Double.MAX_VALUE;

intbestCategoryId=-1;

for(inti=0;i<resultVector.size();i++){

Elemente1=resultVector.getElement(i);

intcategoryId=e1.index();

doublescore=e1.get();

if(score>bestScore){

bestScore=score;

bestCategoryId=categoryId;

}

System.out.print(""+labels.get(categoryId)+":"+score);

}

System.out.println("=>"+labels.get(bestCategoryId));

}

}

Now,putallthesecodesunderoneclassandcreatethe.jarfileofthisclass.Wewillusethis.jarfiletotestournewe-mails.

TestingtheprogramTotesttheprogram,performthefollowingsteps:

1. Createafoldernamedassassinmodeltestinthelocaldirectory,asfollows:

mkdir/tmp/assassinmodeltest

2. Tousethismodel,getthefollowingfilesfromhdfsto/tmp/assassinmodeltest:

Fortheearliercreatedmodel,usethefollowingcommand:

hadoopfs–get/user/hue/prodmodel/tmp/assassinmodeltest

Forlabelindex,usethefollowingcommand:

hadoopfs–get/user/hue/prodlabelindex/tmp/assassinmodeltest

Fordf-countsfromtheassassinvecfolder(changethenameofthepart-00000filetodf-count),usethefollowingcommands:

hadoopfs–get/user/hue/assassinvec/df-count

/tmp/assassinmodeltest

dictionary.file-0fromthesameassassinvecfolder

hadoopfs–get/user/hue/assassinvec/dictionary.file-0

/tmp/assassinmodeltest

3. Under/tmp/assassinmodeltest,createafilewiththemessageshowninthefollowingscreenshot:

4. Now,runtheprogramusingthefollowingcommand:

Java–cp/tmp/assassinmodeltest/spamclassifier.jar:/usr/lib/mahout/*

com.packt.spamfilter.TestClassifier/tmp/assassinmodeltest

/tmp/assassinmodeltest/prodlabelindex

/tmp/assassinmodeltest/dictionary.file-0/tmp/assassinmodeltest/df-

count/tmp/assassinmodeltest/testemail

5. Now,updatetheteste-mailfilewiththemessageshowninthefollowingscreenshot:

6. Runtheprogramagainusingthesamecommandasgiveninstep4andviewtheresultasfollows:

Now,wehaveaprogramreadythatcanuseourclassifiermodelandpredicttheunknownitems.Let’smoveontooursecondusecase.

SecondusecaseasanexerciseAsdiscussedatthestartofthischapter,wewillnowworkonasecondusecase,wherewewillpredictthecategoryofanewe-mail.

TheASFe-maildatasetTheApacheSoftwareFoundatione-maildatasetispartitionedbyproject.Thise-maildatasetcanbefoundathttp://aws.amazon.com/datasets/7791434387204566.

Asmallerdatasetcanbefoundathttp://files.grantingersoll.com/ibm.tar.gz.(Refertohttp://lucidworks.com/blog/scaling-mahout/).Usethisdatatoperformthefollowingsteps:

1. Movethisdatatothefolderofyourchoice(/tmp/asfmail)andunzipthefolder:

mkdir/tmp/asfmail

tar–xvfibm.tar

2. Movethedatasettohdfs:

hadoopfs-put/tmp/asfmail/ibm/content/user/hue/asfmail

3. ConvertthemboxfilesintoHadoop’sSequenceFileformatusingMahout’sSequenceFilesFromMailArchivesasfollows:

mahoutorg.apache.mahout.text.SequenceFilesFromMailArchives--charset

"UTF-8"--body--subject--input/user/hue/asfmail/content--output

/user/hue/asfmailout

http://aws.amazon.com/datasets/7791434387204566

http://files.grantingersoll.com/ibm.tar.gz

http://lucidworks.com/blog/scaling-mahout/

4. Convertthesequencefileintosparsevector:

mahoutseq2sparse--input/user/hue/asfmailout--output

/user/hue/asfmailseqsp--norm2--weightTFIDF--namedVector--

maxDFPercent90--minSupport2--analyzerName

org.apache.mahout.text.MailArchivesClusteringAnalyzer

5. Modifythelabels:

mahoutorg.apache.mahout.classifier.email.PrepEmailDriver--input

/user/hue/asfmailseqsp--output/user/hue/asfmailseqsplabel--

maxItemsPerLabel1000

Now,thenextthreestepsaresimilartotheonesweperformedearlier:

1. Splitthedatasetintotrainingandtestdatasetsusingthefollowingcommand:

mahoutsplit--input/user/hue/asfmailseqsplabel--trainingOutput

/user/hue/asfmailtrain--testOutput/user/hue/asfmailtest--

randomSelectionPct20--overwrite--sequenceFiles

2. Trainthemodelusingthetrainingdatasetasfollows:

mahouttrainnb-i/user/hue/asfmailtrain-o/user/hue/asfmailmodel-

extractLabels--labelIndex/user/hue/asfmaillabels

3. Testthemodelusingthetestdataset:

mahouttestnb-i/user/hue/asfmailtest-m/user/hue/asfmailmodel--

labelIndex/user/hue/asfmaillabels

Asyoumayhavenoticed,allthestepsareexactlyidenticaltotheonesweperformedearlier.Hereby,Ileavethistopicasanexerciseforyoutocreateyourownclassifiersystemusingthismodel.Youcanusehintsasprovidedforthespamfilterclassifier.Wenowmoveourdiscussiontotuningourclassifier.Let’stakeabriefoverviewofthebestpracticesinthisarea.

ClassifierstuningWealreadydiscussedclassifiers’evaluationtechniquesinChapter1,ClassificationinDataAnalysis.Justasareminder,weevaluateourmodelusingtechniquessuchasconfusionmatrix,entropymatrix,areaundercurve,andsoon.

Fromtheexplanatoryvariables,wecreatethefeaturevector.Tocheckhowaparticularmodelisworking,thesefeaturevectorsneedtobeinvestigated.InMahout,thereisaclassavailableforthis,ModelDissector.Ittakesthefollowingthreeinputs:

Features:Thisclasstakesafeaturevectortouse(destructively)TraceDictionary:ThisclasstakesatracedictionarycontainingvariablesandthelocationsinthefeaturevectorthatareaffectedbythemLearner:Thisclasstakesthemodelthatweareprobingtofindweightsonfeatures

ModelDissectortweaksthefeaturevectorandobserveshowthemodeloutputchanges.Bytakinganaverageofthenumberofexamples,wecandeterminetheeffectofdifferentexplanatoryvariables.

ModelDissectorhasasummarymethod,whichreturnsthemostimportantfeatureswiththeirweights,mostimportantcategory,andthetopfewcategoriesthattheyaffect.

TheoutputofModelDissectorishelpfulintroubleshootingproblemsinawronglycreatedmodel.

Moredetailsforthecodecanbefoundathttps://github.com/apache/mahout/blob/master/mrlegacy/src/main/java/org/apache/mahout/classifier/sgd/ModelDissector.java

Whileimprovingtheoutputoftheclassifier,oneshouldtakecarewithtwocommonlyoccurringproblems:targetleak,andbrokenfeatureextraction.

Ifthemodelisshowingresultsthataretoogoodtobetrueoranoutputbeyondexpectations,wecouldhaveaproblemwithtargetleak.Thiserrorcomesonceinformationfromthetargetvariableisincludedintheexplanatoryvariables,whichareusedtotraintheclassifier.Inthisinstance,theclassifierwillworktoowellforthetestdataset.

Ontheotherhand,brokenfeatureextractionoccurswhenfeatureextractionisbroken.Thistypeofclassifiershowstheoppositeresultfromthetargetleakclassifiers.Here,themodelprovidesresultspoorerthanexpected.

Totunetheclassifier,wecanusenewexplanatoryvariables,transformationsofexplanatoryvariables,andcanalsoeliminatesomeofthevariables.Weshouldalsotrydifferentlearningalgorithmstocreatethemodelandchooseanalgorithm,whichisgoodinperformance,trainingtime,andspeed.

MoredetailsontuningcanbefoundinChapter16,DeployingaclassifierinthebookMahoutinAction.

https://github.com/apache/mahout/blob/master/mrlegacy/src/main/java/org/apache/mahout/classifier/sgd/ModelDissector.java

SummaryInthischapter,wediscussedcreatingourownproductionreadyclassifiermodel.Wetookuptwousecaseshere,oneforane-mailspamfilterandtheotherforclassifyingthee-mailaspertheprojects.WeuseddatasetsforApacheSpamAssassinforthee-mailfilterandASFforthee-mailclassifier.

Wealsosawhowtoincreasetheperformanceofyourmodel.

SoyouarenowreadytoimplementclassifiersusingApacheMahoutforyourownrealworldusecases.Happylearning!

IndexA

algorithms,classificationLogisticregression/ClassificationalgorithmsStochasticGradientDescent(SGD)/ClassificationalgorithmsNaïveBayesclassification/ClassificationalgorithmsHiddenMarkovModel(HMM)/Classificationalgorithmsrandomforest/ClassificationalgorithmsMulti-layerperceptron(MLP)/Classificationalgorithms

ApacheSpamAssassinproject/Spame-maildatasetApacheSpark

about/ApacheSparkSparkSQL/ApacheSparkMLib/ApacheSparkGraphX/ApacheSparkSparkstreaming/ApacheSpark

ASFe-maildatasetabout/TheASFe-maildatasetURL/TheASFe-maildataset

Assassindatasetused,forcreatingmodel/CreatingthemodelusingtheAssassindataset

AUC(areaundertheROCcurve)/AreaundertheROCcurveaxons/Neuralnetworkandneurons

Bbackpropagation/MultilayerPerceptronBagofwords/UnderstandingthetermsusedintextclassificationBaumWelchTrainerclass/UsingMahoutfortheHiddenMarkovModelBayesrule

about/IntroducingconditionalprobabilityandtheBayesrulebindingstack

URL/MahoutScalaandSparkbindings

CChi-squaredAutomaticInteractionDetector(CHAID)/Decisiontreeclassification

about/Introducingtheclassification,IntroducingApacheMahoutapplication/Applicationoftheclassificationsystemsystem,working/Workingoftheclassificationsystemalgorithms/Classificationalgorithms

ClassificationandRegressionTree(CART)/Decisiontreeclassifier

trainingdataset/Workingoftheclassificationsystemtestdataset/Workingoftheclassificationsystemmodel/Workingoftheclassificationsystembuilding/Workingoftheclassificationsystem

classifiermodelusing,programfor/Programtouseaclassifiermodel

classifierstuning/Classifierstuning

clusteringabout/IntroducingApacheMahout

conditionalprobabilityabout/IntroducingconditionalprobabilityandtheBayesrule

confusionmatrixabout/TheconfusionmatrixAccuracy/TheconfusionmatrixPrecisionorpositivepredictivevalue/TheconfusionmatrixNegativepredictivevalue/TheconfusionmatrixSensitivity/truepositiverate/recall/TheconfusionmatrixSpecificity/TheconfusionmatrixF1score/Theconfusionmatrix

costfunction,linearregressionabout/Costfunction

DDARPA‘98/UsingMahoutforRandomforestdataanalysis

classification/Introducingtheclassificationdecisiontree

about/Decisiontreedendrites/Neuralnetworkandneuronsdependentvariable/Logisticregressiondeterministicpatterns/Deterministicandnondeterministicpatternsdevelopmentenvironment

settingup,Eclipseused/SettingupadevelopmentenvironmentusingEclipsedimensionalreduction

about/IntroducingApacheMahoutDomainSpecificLanguage(DSL)/Mahoutnewchanges

EEclipse

used,forbuildingdevelopmentenvironment/SettingupadevelopmentenvironmentusingEclipse

emissionmatrix,HMM/IntroducingtheHiddenMarkovModelEntropymatrix

about/Theentropymatrixexplanatoryvariable/Logisticregressionexplanatoryvariables

about/Workingoftheclassificationsystem

Ggradientdescent

about/Gradientdescentsigmoidfunction/Logisticregressionlogisticfunction/Logisticregression

GraphX/ApacheSpark

HH2Oplatform

integration/H2OplatformintegrationURL/H2Oplatformintegration

HadoopURL/IntroducingApacheMahout,InstallingMahout

hiddenlayer,MLPnetwork/MultilayerPerceptronhiddenlayers,MLPnetwork/MultilayerPerceptronHiddenMarkovModel(HMM)/Classificationalgorithmshiddenstates,HMM/IntroducingtheHiddenMarkovModelHMM

about/IntroducingtheHiddenMarkovModelproperties/IntroducingtheHiddenMarkovModelstatevector/IntroducingtheHiddenMarkovModeltransitionmatrix/IntroducingtheHiddenMarkovModelemissionmatrix/IntroducingtheHiddenMarkovModelhiddenstates/IntroducingtheHiddenMarkovModelobservablestate/IntroducingtheHiddenMarkovModelMahoutused/UsingMahoutfortheHiddenMarkovModelModelclass/UsingMahoutfortheHiddenMarkovModelHmmTrainerclass/UsingMahoutfortheHiddenMarkovModelHmmEvaluatorclass/UsingMahoutfortheHiddenMarkovModelHmmAlgorithmsclass/UsingMahoutfortheHiddenMarkovModelHmmUtilsclass/UsingMahoutfortheHiddenMarkovModelRandomSequencerGenerator/UsingMahoutfortheHiddenMarkovModelBaumWelchTrainerclass/UsingMahoutfortheHiddenMarkovModelViterbiEvaluatorclass/UsingMahoutfortheHiddenMarkovModelinputcommand/UsingMahoutfortheHiddenMarkovModeloutputcommand/UsingMahoutfortheHiddenMarkovModelmodelcommand/UsingMahoutfortheHiddenMarkovModellikelihoodcommand/UsingMahoutfortheHiddenMarkovModel

HMM,issuesevaluation/IntroducingtheHiddenMarkovModeldecoding/IntroducingtheHiddenMarkovModellearning/IntroducingtheHiddenMarkovModel

HmmAlgorithmsclass/UsingMahoutfortheHiddenMarkovModelHmmEvaluatorclass/UsingMahoutfortheHiddenMarkovModelHMMModelclass/UsingMahoutfortheHiddenMarkovModelHmmTrainerclass/UsingMahoutfortheHiddenMarkovModelHmmUtilsclass/UsingMahoutfortheHiddenMarkovModelHortonworksSandbox

URL/SettingupMahoutforaWindowsuser

IiInitialstatevector,Markovprocess/TheMarkovprocessindependentvariable/Logisticregressioninputlayer,MLPnetwork/MultilayerPerceptronirisdataset

URL/UsingMahoutforMLPIterativeDichotomiser3(ID3)

URL/Decisiontree

JJava

URL/InstallingMahout

Llabels

about/WorkingoftheclassificationsystemLatentDirichletAllocation(LDA)/IntroducingApacheMahoutlinearregression

about/Understandinglinearregressioncostfunction/Costfunctiongradientdescent/Gradientdescent

logisticfunction/Logisticregressionlogisticregression/Classificationalgorithms

about/LogisticregressionMahout,usingfor/UsingMahoutforlogisticregressiondataset/UsingMahoutforlogisticregressiontrainingandtestdata,preparing/UsingMahoutforlogisticregressionmodel,training/UsingMahoutforlogisticregressiontrainlogistic/UsingMahoutforlogisticregressioninput/UsingMahoutforlogisticregressionoutput/UsingMahoutforlogisticregressiontarget/UsingMahoutforlogisticregressioncategories/UsingMahoutforlogisticregressionpredictors/UsingMahoutforlogisticregressiontypes/UsingMahoutforlogisticregressionfeatures/UsingMahoutforlogisticregressionpasses/UsingMahoutforlogisticregressionrate/UsingMahoutforlogisticregressionrunlogistic/UsingMahoutforlogisticregressionmodel/UsingMahoutforlogisticregressionauc/UsingMahoutforlogisticregressionconfusion/UsingMahoutforlogisticregression

MM2Eclipse

URL/InstallingMavenMahout

about/IntroducingApacheMahoutusecases/IntroducingApacheMahoutfeatures/ReasonsforMahoutbeingagoodchoiceforclassificationinstalling/InstallingMahoutprerequisites/InstallingMahoutbuildingfromsource,Mavenused/BuildingMahoutfromsourceusingMavenMaven,installing/InstallingMavencode,building/BuildingMahoutcodedistributionfile,URL/BuildingMahoutcode,SettingupadevelopmentenvironmentusingEclipsesettingup,forWindowsuser/SettingupMahoutforaWindowsuserused,forlogisticregression/UsingMahoutforlogisticregressionNaïveBayesalgorithm/UsingtheNaïveBayesalgorithminApacheMahoutusing,forHMM/UsingMahoutfortheHiddenMarkovModelusing,forRandomforestalgorithm/UsingMahoutforRandomforestRandomforestalgorithm,implementing/StepstousetheRandomforestalgorithminMahoutMLP,implementing/MLPimplementationinMahoutusing,forMLP/UsingMahoutforMLPMLPalgorithm,using/StepstousetheMLPalgorithminMahoutupdations/MahoutnewchangesScalabindings/MahoutScalaandSparkbindingsSparkbindings/MahoutScalaandSparkbindingsSparkshell,using/UsingMahout’sSparkshellH2Oplatform,integration/H2Oplatformintegration

Mahout,algorithmsabout/AlgorithmssupportedinMahoutsequentialalgorithms/AlgorithmssupportedinMahoutparallelalgorithms/AlgorithmssupportedinMahout

Mahout,usecasesrecommendation/IntroducingApacheMahoutclassification/IntroducingApacheMahoutclustering/IntroducingApacheMahoutdimensionalreduction/IntroducingApacheMahouttopicmodeling/IntroducingApacheMahout

MahoutScalabindingsabout/MahoutScalaandSparkbindings

MahoutSparkbindingsabout/MahoutScalaandSparkbindings

Markovprocessabout/TheMarkovprocessstates/TheMarkovprocesstransitionmatrix/TheMarkovprocessTransitionmatrix/TheMarkovprocessInitialstatevector/TheMarkovprocess

Mavenused,forbuildingMahoutfromsource/BuildingMahoutfromsourceusingMaveninstalling/InstallingMavenURL/InstallingMaven

MLib/ApacheSparkMLP

implementing,inMahout/MLPimplementationinMahoutMahoutused/UsingMahoutforMLPirisdataset/UsingMahoutforMLP

MLPalgorithmusing,inMahout/StepstousetheMLPalgorithminMahout

MLPnetworkabout/MultilayerPerceptronhiddenlayers/MultilayerPerceptronbackpropagation/MultilayerPerceptronzerohiddenlayers/MultilayerPerceptroninputlayer/MultilayerPerceptronoutputlayer/MultilayerPerceptronhiddenlayer/MultilayerPerceptronnumberofneuronsorhiddenunits/MultilayerPerceptron

modelcreating,Assassindatasetused/CreatingthemodelusingtheAssassindatasetclassifiermodel,programforusing/Programtouseaclassifiermodel

model,evaluationconfusionmatrix/TheconfusionmatrixReceiverOperatingCharacteristics(ROC)graph/TheReceiverOperatingCharacteristics(ROC)graphareaundertheROCcurve(AUC)/AreaundertheROCcurveEntropymatrix/Theentropymatrix

model,issuesoverfitting/Workingoftheclassificationsystemunderfitting/Workingoftheclassificationsystem

ModelDissectorFeaturesclass/ClassifierstuningTraceDictionaryclass/ClassifierstuningLearnerclass/Classifierstuningabout/Classifierstuning

Multi-layerperceptron(MLP)/Classificationalgorithms

NNaïveBayesalgorithm

about/UnderstandingtheNaïveBayesalgorithminApacheMahout/UsingtheNaïveBayesalgorithminApacheMahout

NaïveBayesclassification/Classificationalgorithmsneuralnetwork

about/Neuralnetworkandneuronsneurons

about/NeuralnetworkandneuronsURL/Neuralnetworkandneurons

nondeterministicpatterns/DeterministicandnondeterministicpatternsNSL-KDDdataset

URL/UsingMahoutforRandomforest

Oobservablestate,HMM/IntroducingtheHiddenMarkovModeloutlierdetection

about/Workingoftheclassificationsystemoutputlayer,MLPnetwork/MultilayerPerceptronoverfitting,model

issues/Workingoftheclassificationsystem

Pparallelalgorithms/AlgorithmssupportedinMahoutprogram

testing/Testingtheprogrampruning/Decisiontree

Rrandomforest/ClassificationalgorithmsRandomforestalgorithm

about/RandomforestBiasparameter/RandomforestVarianceparameter/RandomforestMahoutused/UsingMahoutforRandomforestNSL-KDDdataset/UsingMahoutforRandomforestdataset/UsingMahoutforRandomforestimplementing,inMahout/StepstousetheRandomforestalgorithminMahout

RandomSequencerGenerator/UsingMahoutfortheHiddenMarkovModelReceiverOperatingCharacteristics(ROC)graph

about/TheReceiverOperatingCharacteristics(ROC)graphregression

about/Introducingregressionlinearregression/Understandinglinearregression

regressionintercept/Logisticregression

Ssequentialalgorithms/AlgorithmssupportedinMahoutsigmoidfunction/Logisticregressionsoftmaxfunction

URL/Neuralnetworkandneuronsspame-maildatasetclassifier

about/Spame-maildatasetSpark

URL/UsingMahout’sSparkshellbinding,URL/UsingMahout’sSparkshell

Spark-item/ApacheSparkSpark-row/ApacheSparkSparkshell

using/UsingMahout’sSparkshellSparkSQL/ApacheSparkSparkstreaming/ApacheSparkstates,Markovprocess/TheMarkovprocessstatevector,HMM/IntroducingtheHiddenMarkovModelStochasticGradientDescent(SGD)/Classificationalgorithms

about/StochasticGradientDescent

Ttargetvariables

about/WorkingoftheclassificationsystemTermfrequency/Understandingthetermsusedintextclassificationtermfrequency

Stemmingofwords/UnderstandingthetermsusedintextclassificationCasenormalization/UnderstandingthetermsusedintextclassificationStopwordremoval/UnderstandingthetermsusedintextclassificationInversedocumentfrequency/UnderstandingthetermsusedintextclassificationTermfrequencyandinversetermfrequency/Understandingthetermsusedintextclassification

textclassificationabout/Understandingthetermsusedintextclassification

topicmodelingabout/IntroducingApacheMahout

transitionmatrix,HMM/IntroducingtheHiddenMarkovModeltransitionmatrix,Markovprocess/TheMarkovprocess

Uunderfitting,model

issues/Workingoftheclassificationsystem

Vvectors

about/UnderstandingthetermsusedintextclassificationViterbiEvaluatorclass/UsingMahoutfortheHiddenMarkovModel

WWindows

user,Mahoutsettingupfor/SettingupMahoutforaWindowsuserWisconsinDiagnosticBreastCancer(WDBC)dataset

URL/UsingMahoutforlogisticregression

Learning Apache Mahout Classification - DropPDF1.droppdf.com/files/0ZaSP/learning-apache-mahout-classification-by... · Learning Apache Mahout Classification Credits ... , with PDF

Documents