1

arXiv:2108.02497v3[cs.LG]9Feb2023

Howtoavoidmachinelearningpitfalls:aguideforacademicresearchers

MichaelA.Lones*

Abstract

Thisdocumentisaconciseoutlineofsomeofthecommonmistakesthatoccurwhen

usingmachinelearning,andwhatcanbedonetoavoidthem.Whilstitshouldbeaccessibletoanyonewithabasicunderstandingofmachinelearningtechniques,itwasoriginallywrittenforresearchstudents,andfocusesonissuesthatareofpartic-ularconcernwithinacademicresearch,suchastheneedtodorigorouscomparisonsandreachvalidconclusions.Itcovers?vestagesofthemachinelearningprocess:whattodobeforemodelbuilding,howtoreliablybuildmodels,howtorobustlyevaluatemodels,howtocomparemodelsfairly,andhowtoreportresults.

1Introduction

It’seasytomakemistakeswhenapplyingmachinelearning(ML),andthesemistakescanresultinMLmodelsthatfailtoworkasexpectedwhenappliedtodatanotseenduringtrainingandtesting[

Liaoetal.

,

2021

].Thisisaproblemforpractitioners,sinceitleadstothefailureofMLprojects.However,itisalsoaproblemforsociety,sinceiterodestrustinthe?ndingsandproductsofML[

Gibney

,

2022

].Thisguideaimstohelpnewcomersavoidsomeofthesemistakes.It’swrittenbyanacademic,andfocusesonlessonslearntwhilstdoingMLresearchinacademia.Whilstprimarilyaimedatstudentsandscienti?cresearchers,itshouldbeaccessibletoanyonegettingstartedinML,andonlyassumesabasicknowledgeofMLtechniques.However,unlikesimilarguidesaimedatamoregeneralaudience,itincludestopicsthatareofaparticularconcerntoacademia,suchastheneedtorigorouslyevaluateandcomparemodelsinordertogetworkpublished.Tomakeitmorereadable,theguidanceiswritteninformally,inaDosandDon’tsstyle.It’snotintendedtobeexhaustive,andreferences(withpublicly-accessibleURLswhereavailable)areprovidedforfurtherreading.Sinceitdoesn’tcoverissuesspeci?ctoparticularacademicsubjects,it’srecommendedyoualsoconsultsubject-speci?cguidancewhereavailable(e.g.

Stevensetal.

[

2020]

formedicine).Feedbackiswelcome,anditisexpectedthatthisdocumentwillevolveovertime.Forthisreason,ifyouciteit,pleaseincludethearXivversionnumber(currentlyv3).

*SchoolofMathematicalandComputerSciences,Heriot-WattUniversity,Edinburgh,Scotland,UK,Email:

m.lones@hw.ac.uk

,Web:

http://www.macs.hw.ac.uk/~ml355

.

2

Contents

1Introduction

1

2Beforeyoustarttobuildmodels

3

2.1Dotakethetimetounderstandyourdata

3

2.2Don’tlookatallyourdata

3

2.3Domakesureyouhaveenoughdata

3

2.4Dotalktodomainexperts

4

2.5Dosurveytheliterature

4

2.6Dothinkabouthowyourmodelwillbedeployed

5

3Howtoreliablybuildmodels

5

3.1Don’tallowtestdatatoleakintothetrainingprocess

5

3.2Dotryoutarangeofdi?erentmodels

6

3.3Don’tuseinappropriatemodels

7

3.4Dokeepupwithrecentdevelopmentsindeeplearning

8

3.5Don’tassumedeeplearningwillbethebestapproach

8

3.6Dooptimiseyourmodel’shyperparameters

9

3.7Dobecarefulwhereyouoptimisehyperparametersandselectfeatures

9

3.8Doavoidlearningspuriouscorrelations

11

4Howtorobustlyevaluatemodels

11

4.1Douseanappropriatetestset

11

4.2Don’tdodataaugmentationbeforesplittingyourdata

12

4.3Douseavalidationset

12

4.4Doevaluateamodelmultipletimes

12

4.5Dosavesomedatatoevaluateyour?nalmodelinstance

14

4.6Don’tuseaccuracywithimbalanceddatasets

14

4.7Don’tignoretemporaldependenciesintimeseriesdata

15

5Howtocomparemodelsfairly

16

5.1Don’tassumeabiggernumbermeansabettermodel

16

5.2Dousestatisticaltestswhencomparingmodels

16

5.3Docorrectformultiplecomparisons

17

5.4Don’talwaysbelieveresultsfromcommunitybenchmarks

17

5.5Doconsidercombinationsofmodels

17

6Howtoreportyourresults

18

6.1Dobetransparent

18

6.2Doreportperformanceinmultipleways

19

6.3Don’tgeneralisebeyondthedata

19

6.4Dobecarefulwhenreportingstatisticalsigni?cance

19

6.5Dolookatyourmodels

20

7Finalthoughts

20

8Acknowledgements

21

9Changes

21

3

2Beforeyoustarttobuildmodels

It’snormaltowanttorushintotrainingandevaluatingmodels,butit’simportanttotakethetimetothinkaboutthegoalsofaproject,tofullyunderstandthedatathatwillbeusedtosupportthesegoals,toconsideranylimitationsofthedatathatneedtobeaddressed,andtounderstandwhat’salreadybeendoneinyour?eld.Ifyoudon’tdothesethings,thenyoumayendupwithresultsthatarehardtopublish,ormodelsthatarenotappropriatefortheirintendedpurpose.

2.1Dotakethetimetounderstandyourdata

Eventuallyyouwillwanttopublishyourwork.Thisisaloteasiertodoifyourdataisfromareliablesource,hasbeencollectedusingareliablemethodology,andisofgoodquality.Forinstance,ifyouareusingdatacollectedfromaninternetresource,makesureyouknowwhereitcamefrom.Isitdescribedinapaper?Ifso,takealookatthepaper;makesureitwaspublishedsomewherereputable,andcheckwhethertheauthorsmentionanylimitationsofthedata.Donotassumethat,becauseadatasethasbeenusedbyanumberofpapers,itisofgoodquality—sometimesdataisusedjustbecauseitiseasytogetholdof,andsomewidelyuseddatasetsareknowntohavesigni?cantlimitations(see

Paulladaetal.

[

2020

]foradiscussionofthis).Ifyoutrainyourmodelusingbaddata,thenyouwillmostlikelygenerateabadmodel:aprocessknownasgarbageingarbageout.So,alwaysbeginbymakingsureyourdatamakessense.Dosomeexploratorydataanalysis(see

Cox

[

2017

]forsuggestions).Lookformissingorinconsistentrecords.Itismucheasiertodothisnow,beforeyoutrainamodel,ratherthanlater,whenyou’retryingtoexplaintoreviewerswhyyouusedbaddata.

2.2Don’tlookatallyourdata

Asyoulookatdata,itisquitelikelythatyouwillspotpatternsandmakeinsightsthatguideyourmodelling.Thisisanothergoodreasontolookatdata.However,itisimportantthatyoudonotmakeuntestableassumptionsthatwilllaterfeedintoyourmodel.The“untestable”bitisimportanthere;it’s?netomakeassumptions,buttheseshouldonlyfeedintothetrainingofthemodel,notthetesting.So,toensurethisisthecase,youshouldavoidlookingcloselyatanytestdataintheinitialexploratoryanalysisstage.Otherwiseyoumight,consciouslyorunconsciously,makeassumptionsthatlimitthegeneralityofyourmodelinanuntestableway.ThisisathemeIwillreturntoseveraltimes,sincetheleakageofinformationfromthetestsetintothetrainingprocessisacommonreasonwhyMLmodelsfailtogeneralise.See

Don’tallowtestdatatoleakinto

thetrainingprocess

formoreonthis.

2.3Domakesureyouhaveenoughdata

Ifyoudon’thaveenoughdata,thenitmaynotbepossibletotrainamodelthatgener-alises.Workingoutwhetherthisisthecasecanbechallenging,andmaynotbeevidentuntilyoustartbuildingmodels:italldependsonthesignaltonoiseratiointhedataset.

4

Ifthesignalisstrong,thenyoucangetawaywithlessdata;ifit’sweak,thenyouneedmoredata.Ifyoucan’tgetmoredata—andthisisacommonissueinmanyresearch?elds—thenyoucanmakebetteruseofexistingdatabyusingcross-validation(see

Doevaluateamodelmultipletimes

).Youcanalsousedataaugmentationtechniques(e.g.see

Wongetal.

[

2016

]and

ShortenandKhoshgoftaar

[

2019

];fortimeseriesdata,see

IwanaandUchida

[

2021

]),andthesecanbequitee?ectiveforboostingsmalldatasets,though

Don’tdodataaugmentationbeforesplittingyourdata

.Dataaugmentationisalsousefulinsituationswhereyouhavelimiteddataincertainpartsofyourdataset,e.g.inclassi?cationproblemswhereyouhavelesssamplesinsomeclassesthanothers,asituationknownasclassimbalance.See

Haixiangetal.

[

2017

]forareviewofmethodsfordealingwiththis;alsosee

Don’tuseaccuracywithimbalanceddatasets

.Anotheroptionfordealingwithsmalldatasetsistousetransferlearning(see

Dokeepupwith

recentdevelopmentsindeeplearning

).However,ifyouhavelimiteddata,thenit’slikelythatyouwillalsohavetolimitthecomplexityoftheMLmodelsyouuse,sincemodelswithmanyparameters,likedeepneuralnetworks,caneasilyover?tsmalldatasets(see

Don’tassumedeeplearningwillbethebestapproach

).Eitherway,it’simportanttoidentifythisissueearlyon,andcomeupwithasuitablestrategytomitigateit.

2.4Dotalktodomainexperts

Domainexpertscanbeveryvaluable.Theycanhelpyoutounderstandwhichproblemsareusefultosolve,theycanhelpyouchoosethemostappropriatefeaturesetandMLmodeltouse,andtheycanhelpyoupublishtothemostappropriateaudience.Failingtoconsidertheopinionofdomainexpertscanleadtoprojectswhichdon’tsolveusefulproblems,orwhichsolveusefulproblemsininappropriateways.AnexampleofthelatterisusinganopaqueMLmodeltosolveaproblemwherethereisastrongneedtounderstandhowthemodelreachesanoutcome,e.g.inmakingmedicalor?nancialdecisions(see

Rudin

[

2019

]).Atthebeginningofaproject,domainexpertscanhelpyoutounderstandthedata,andpointyoutowardsfeaturesthatarelikelytobepredictive.Attheendofaproject,theycanhelpyoutopublishindomain-speci?cjournals,andhencereachanaudiencethatismostlikelytobene?tfromyourresearch.

2.5Dosurveytheliterature

You’reprobablynotthe?rstpersontothrowMLataparticularproblemdomain,soit’simportanttounderstandwhathasandhasn’tbeendonepreviously.Otherpeoplehavingworkedonthesameproblemisn’tabadthing;academicprogressistypicallyaniterativeprocess,witheachstudyprovidinginformationthatcanguidethenext.Itmaybediscouragingto?ndthatsomeonehasalreadyexploredyourgreatidea,buttheymostlikelyleftplentyofavenuesofinvestigationstillopen,andtheirpreviousworkcanbeusedasjusti?cationforyourwork.Toignorepreviousstudiesistopotentiallymissoutonvaluableinformation.Forexample,someonemayhavetriedyourproposedapproachbeforeandfoundfundamentalreasonswhyitwon’twork(andthereforesavedyouafewyearsoffrustration),ortheymayhavepartiallysolvedtheprobleminawaythatyou

canbuildon.So,it’simportanttodoaliteraturereviewbeforeyoustartwork;leavingittoolatemaymeanthatyouareleftscramblingtoexplainwhyyouarecoveringthesamegroundornotbuildingonexistingknowledgewhenyoucometowriteapaper.

2.6Dothinkabouthowyourmodelwillbedeployed

WhydoyouwanttobuildanMLmodel?Thisisanimportantquestion,andtheanswershouldin?uencetheprocessyouusetodevelopyourmodel.Manyacademicstudiesarejustthat—studies—andnotreallyintendedtoproducemodelsthatwillbeusedintherealworld.Thisisfairenough,sincetheprocessofbuildingandanalysingmodelscanitselfgiveveryusefulinsightsintoaproblem.However,formanyacademicstudies,theeventualgoalistoproduceanMLmodelthatcanbedeployedinarealworldsituation.Ifthisisthecase,thenit’sworththinkingearlyonabouthowitisgoingtobedeployed.Forinstance,ifit’sgoingtobedeployedinaresource-limitedenvironment,suchasasensororarobot,thismayplacelimitationsonthecomplexityofthemodel.Iftherearetimeconstraints,e.g.aclassi?cationofasignalisrequiredwithinmilliseconds,thenthisalsoneedstobetakenintoaccountwhenselectingamodel.Anotherconsiderationishowthemodelisgoingtobetiedintothebroadersoftwaresystemwithinwhichitisdeployed;thisprocedureisoftenfarfromsimple(see

Sculley

etal.

[

2015

]).However,emergingapproachessuchasMLOpsaimtoaddresssomeofthedi?culties;see

Tamburri

[

2020

]forareview,and

Shankaretal.

[

2022

]foradiscussionofcommonchallengeswhenoperationalisingMLmodels.

3Howtoreliablybuildmodels

BuildingmodelsisoneofthemoreenjoyablepartsofML.WithmodernMLframeworks,it’seasytothrowallmannerofapproachesatyourdataandseewhatsticks.However,thiscanleadtoadisorganisedmessofexperimentsthat’shardtojustifyandhardtowriteup.So,it’simportanttoapproachmodelbuildinginanorganisedmanner,makingsureyouusedatacorrectly,andputtingadequateconsiderationintothechoiceofmodels.

3.1Don’tallowtestdatatoleakintothetrainingprocess

It’sessentialtohavedatathatyoucanusetomeasurehowwellyourmodelgeneralises.Acommonproblemisallowinginformationaboutthisdatatoleakintothecon?guration,trainingorselectionofmodels(seeFigure

1

).Whenthishappens,thedatanolongerprovidesareliablemeasureofgenerality,andthisisacommonreasonwhypublishedMLmodelsoftenfailtogeneralisetorealworlddata.Thereareanumberofwaysthatinformationcanleakfromatestset.Someoftheseseemquiteinnocuous.Forinstance,duringdatapreparation,usinginformationaboutthemeansandrangesofvariableswithinthewholedatasettocarryoutvariablescaling—inordertopreventinformationleakage,thiskindofthingshouldonlybedonewiththetrainingdata.Othercommonexamplesofinformationleakagearecarryingoutfeatureselectionbeforepartitioningthedata(see

Dobecarefulwhereyouoptimisehyperparametersandselect

5

Figure1:See

Don’tallowtestdatatoleakintothetrainingprocess

.[left]Howthingsshouldbe,withthetrainingsetusedtotrainthemodel,andthetestsetusedtomeasureitsgenerality.[right]Whenthere’sadataleak,thetestsetcanimplicitlybecomepartofthetrainingprocess,meaningthatitnolongerprovidesarealiablemeasureofgenerality.

features

),usingthesametestdatatoevaluatethegeneralityofmultiplemodels(see

Douseavalidationset

and

Don’talwaysbelieveresultsfromcommunitybenchmarks

),andapplyingdataaugmentationbeforesplittingo?thetestdata(see

Don’tdodata

augmentationbeforesplittingyourdata

).Thebestthingyoucandotopreventtheseissuesistopartitiono?asubsetofyourdatarightatthestartofyourproject,andonlyusethisindependenttestsetoncetomeasurethegeneralityofasinglemodelattheendoftheproject(see

Dosavesomedatatoevaluateyour?nalmodelinstance

).Beparticularlycarefulifyou’reworkingwithtimeseriesdata,sincerandomsplitsofthedatacaneasilycauseleakageandover?tting—see

Don’tignoretemporaldependencies

intimeseriesdata

formoreonthis.Forabroaderdiscussionofdataleakage,see

Kapoor

andNarayanan

[

2022

].

3.2Dotryoutarangeofdi?erentmodels

Generallyspeaking,there’snosuchthingasasinglebestMLmodel.Infact,there’saproofofthis,intheformoftheNoFreeLunchtheorem,whichshowsthatnoMLapproachisanybetterthananyotherwhenconsideredovereverypossibleproblem[

Wolpert

,

2002

].So,yourjobisto?ndtheMLmodelthatworkswellforyourparticularproblem.Thereissomeguidanceonthis.Forexample,youcanconsidertheinductivebiasesofMLmodels;thatis,thekindofrelationshipstheyarecapableofmodelling.Forinstance,linearmodels,suchaslinearregressionandlogisticregression,areagoodchoiceifyouknowtherearenoimportantnon-linearrelationshipsbetweenthefeaturesinyourdata,butabadchoiceotherwise.Goodqualityresearchoncloselyrelatedproblemsmayalsobeabletopointyoutowardsmodelsthatworkparticularlywell.However,alotofthetimeyou’restillleftwithquiteafewchoices,andtheonlywaytoworkoutwhichmodelisbestistotrythemall.Fortunately,modernMLlibrariesinPython(e.g.scikit-learn[

Varoquauxetal.

,

2015

]),R(e.g.caret[

Kuhn

,

2015]

),Julia(e.g.MLJ[

Blaometal.

,

2020

])etc.allowyoutotryoutmultiplemodelswithonlysmallchangestoyourcode,sothere’snoreasonnottotrythemalloutand?ndoutforyourselfwhichoneworksbest.However,

Don’tuseinappropriatemodels

,and

Douse

6

7

Figure2:See

Dokeepupwithrecentdevelopmentsindeeplearning

.Aroughhistoryofneuralnetworksanddeeplearning,showingwhatIconsidertobethemilestonesintheirdevelopment.Forafarmorethoroughandaccurateaccountofthe?eld’shistoricaldevelopment,takealookat

Schmidhuber

[

2015

].

avalidationset

,ratherthanthetestset,toevaluatethem.Whencomparingmodels,

Dooptimiseyourmodel’shyperparameters

and

Doevaluateamodelmultipletimes

tomakesureyou’regivingthemallafairchance,and

Docorrectformultiplecomparisons

whenyoupublishyourresults.

3.3Don’tuseinappropriatemodels

Byloweringthebarriertoimplementation,modernMLlibrariesalsomakeiteasytoapplyinappropriatemodelstoyourdata.This,inturn,couldlookbadwhenyoutrytopublishyourresults.Asimpleexampleofthisisapplyingmodelsthatexpectcategoricalfeaturestoadatasetcontainingnumericalfeatures,orviceversa.SomeMLlibrariesallowyoutodothis,butitmayresultinapoormodelduetolossofinformation.Ifyoureallywanttousesuchamodel,thenyoushouldtransformthefeatures?rst;therearevariouswaysofdoingthis,rangingfromsimpleone-hotencodingstocomplexlearnedembeddings.Otherexamplesofinappropriatemodelchoiceincludeusingaclassi?cationmodelwherearegressionmodelwouldmakemoresense(orviceversa),attemptingtoapplyamodelthatassumesnodependenciesbetweenvariablestotimeseriesdata,orusingamodelthatisunnecessarilycomplex(see

Don’tassumedeeplearningwillbethe

bestapproach

).Also,ifyou’replanningtouseyourmodelinpractice,

Dothinkabout

howyourmodelwillbedeployed

,anddon’tusemodelsthataren’tappropriateforyourusecase.

8

3.4Dokeepupwithrecentdevelopmentsindeeplearning

Machinelearningisafast-moving?eld,andit’seasytofallbehindthecurveanduseapproachesthatotherpeopleconsidertobeoutmoded.Nowhereisthismorethecasethanindeeplearning.So,whilstdeeplearningmaynotalwaysbethebestsolution(see

Don’tassumedeeplearningwillbethebestapproach

),ifyouaregoingtousedeeplearning,thenit’sadvisabletotryandkeepupwithrecentdevelopments.Togivesomeinsightintothis,Figure

2

summarisessomeoftheimportantdevelopmentsovertheyears.Multilayerperceptrons(MLP)andrecurrentneuralnetworks(particularlyLSTM)havebeenpopularforsometime,butareincreasinglybeingreplacedbynewermodelssuchasconvolutionalneuralnetworks(CNN)andtransformers.CNNs(see

Lietal.

[

2021

]forareview)arenowthego-tomodelformanytasks,andcanbeappliedtobothimagedataandnon-imagedata.Beyondtheuseofconvolutionallayers,someofthemainmilestoneswhichledtothesuccessofCNNsincludetheuseofrecti?edlinearunits(ReLU),theadoptionofmodernoptimisers(notablyAdamanditsvariants)andthewidespreaduseofregularisation,especiallydropoutlayersandbatchnormalisation—sogiveseriousconsiderationtoincludingtheseinyourmodels.Anotherimportantgroupofcontemporarymodelsaretransformers(see

Linetal.

[

2022

]forareview).Thesearegraduallyreplacingrecurrentneuralnetworksasthego-tomodelforprocessingsequentialdata,andareincreasinglybeingappliedtootherdatatypestoo,suchasimages[

Khanetal.

,

2022

].AprominentdownsideofbothtransformersanddeepCNNsisthattheyhavemanyparametersandthereforerequirealotofdatatotrainthem.However,anoptionforsmalldatasetsistousetransferlearning,whereamodelispre-trainedonalargegenericdatasetandthen?ne-tunedonthedatasetofinterest[

Hanetal.

,

2021

].Foranextensive,yetaccessible,guidetodeeplearning,see

Zhangetal.

[

2021

].

3.5Don’tassumedeeplearningwillbethebestapproach

Anincreasinglycommonpitfallistoassumethatdeepneuralnetworkswillprovidethebestsolutiontoanyproblem,andconsequentlyfailtotryoutother,possiblymoreappropriate,models.Whilstdeeplearningisgreatforcertaintasks,itisnotgoodateverything;thereareplentyofexamplesofitbeingout-performedby“oldfashioned”machinelearningmodelssuchasrandomforestsandSVMs.See,forinstance,

Grinsztajn

etal.

[

2022

],whoshowthattree-basedmodelsoftenoutperformdeeplearnersontabulardata.Certainkindsofdeepneuralnetworkarchitecturemayalsobeill-suitedtocertainkindsofdata:see,forexample,

Zengetal.

[

2022

],whoarguethattransformersarenotwell-suitedtotimeseriesforecasting.Therearealsotheoreticalreasonswhyanyonekindofmodelwon’talwaysbethebestchoice(see

Dotryoutarangeofdi?erentmodels

).Inparticular,adeepneuralnetworkisunlikelytobeagoodchoiceifyouhavelimiteddata,ifdomainknowledgesuggeststhattheunderlyingpatternisquitesimple,orifthemodelneedstobeinterpretable.Thislastpointisparticularlyworthconsider