Machine Learning & its Classification Techniques

International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-9S3, July 2019

Machine Learning & its Classification Techniques

Atul B.Kathole, Prasad S. Halgaonkar, Ashvini A.Nikhade

Abstract:T RecentT developmentsT inT informationT systemsT asT

wellT asT computerizationT ofT businessT processesT byT

organizationsT haveT ledT toT aT faster,T easierT andT moreT accurateT dataT analysisT andT accuracy.T DataT miningT andT machineT learningT techniquesT haveT beenT usedT increasinglyT inT theT analysisT ofT dataT inT variousT fieldsT rangingT fromT

medicineT toT organization,T educationT andT energyT

applications.T MachineT learningT techniquesT makeT itT

possibleT toT deductT meaningfulT furtherT informationT fromT thoseT dataT processedT byT dataT mining.T SuchT meaningfulT

andT significantT informationT helpsT organizationsT toT

establishT theirT futureT policies.T ThisT studyT appliesT classificationT machineT learningT techniquesT alsoT surveyT ofT

activeT learningT regardingT selectionT methods,T queryT

strategies,T applications.

Keyword: Machine Learning, Data Mining,

Classification Techniques.

I. INTRODUCTION:

MachineT LearningT isT anT approachT orT subsetT ofT

ArtificialT IntelligenceT thatT isT basedT onT theT ideaT thatT

machinesT canT beT givenT accessT toT dataT alongT withT theT

abilityT toT learnT fromT it.T ItT learnsT fromT examplesT andT

experience,T withoutT beingT explicitlyT programmed.T

InsteadT ofT writingT code,T youT feedT dataT toT theT genericT

algorithm,T andT itT buildsT logicT basedT onT theT dataT

given.T InT aT veryT laymanT manner,T MachineT LearningT

(ML)T canT beT explainedT asT automatingT andT improvingT

theT learningT processT ofT computersT basedT onT theirT

experiencesT withoutT beingT actuallyT programmedT i.e.T

withoutT anyT humanT assistance.T TheT processT startsT withT

feedingT goodT qualityT dataT andT thenT trainingT ourT

machinesT (computers)T byT buildingT machineT learningT

modelsT usingT theT dataT andT differentT algorithms.T TheT

choiceT ofT algorithmsT dependsT onT whatT typeT ofT dataT

doT weT haveT andT whatT kindT ofT taskT weT areT tryingT toT

automate.

Revised Manuscript Received on July 25, 2019.

Atul B.Kathole, Faculty,Department of Computer

Engineering, Zeal College of Engineering and Research, Narhe, Pune, India.

Dr.Prasad S. Halgaonkar, Faculty, Department of

Computer Engineering, Zeal College of Engineering and Research, Narhe, Pune, India.

Ashvini A.Nikhade, Faculty, Department of Computer

Engineering, Zeal College of Engineering and Research, Narhe, Pune, India.

[email protected],prasad.halgaonkar@zealeducat ion.com, [email protected]

Data Science and Machine Learning

 DataT ScienceT andT MachineT LearningT goT handT

inT hand.T DataT ScienceT helpsT evaluateT dataT forT

MachineT LearningT algorithms

 DataT scienceT isT theT useT ofT statisticalT methodsT

toT findT patternsT inT theT data.

 StatisticalT machineT learningT usesT theT sameT

mathT andT techniquesT asT dataT science.

 TheseT techniquesT areT integratedT intoT

algorithmsT thatT learnT andT improveT onT theirT

own.

 MachineT LearningT facilitatesT ArtificialT

IntelligenceT asT itT enablesT machinesT toT learnT

fromT theT patternsT inT data.

DifferenceT betweenT TraditionalT ProgrammingT &T

MachineT LearningT

 TraditionalT Programming:T WeT feedT inT

DATAT (Input)T +T PROGRAMT (logic),T runT itT onT

machineT andT getT output.

 MachineT LearningT :T WeT feedT inT

DATA(Input)T +T Output,T runT itT onT machineT

duringT trainingT andT theT machineT createsT itsT ownT

program(logic),T whichT canT beT evaluatedT whileT

testing.

Figure 01:Difference between Traditional Programming & Machine Learning

What does exactly learning means for a computer? AT computerT isT saidT toT beT learningT fromT ExperiencesT

withT respectT toT someT classT ofT Tasks,T ifT itsT

performanceT inT aT givenT TaskT improvesT withT theT

Experience.T AT computerT programT isT saidT toT learnT

fromT experienceT ET withT respectT toT someT classT ofT

tasksT TT andT performanceT measureT P,T ifT itsT

improvesT withT experienceT E Need of Machine Learning

MachineT LearningT isT aT fieldT whichT isT raisedT

outT ofT ArtificialT IntelligenceT (AI).T ApplyingT AI,T weT

wantedT toT buildT betterT andT intelligentT machines.T ButT

exceptT forT fewT mereT tasksT suchT asT findingT theT

shortestT pathT betweenT pointT AT andT B,T weT wereT unableT

toT programT moreT complexT andT constantlyT evolvingT

challenges.

ThereT wasT aT realizationT thatT theT onlyT wayT toT

beT ableT toT achieveT thisT taskT wasT toT letT machineT learnT

fromT itself.T ThisT soundsT similarT toT aT childT learningT

fromT itsT self.T SoT machineT learningT wasT developedT asT

aT newT capabilityT forT computers.T AndT nowT machineT

learningT isT presentT inT soT manyT segmentsT ofT

technology,T thatT weT don’tT evenT realizeT itT whileT usingT

it.T FindingT patternsT inT dataT onT planetT earthT isT possibleT

onlyT forT humanT brains.T TheT dataT beingT veryT massive,T

theT timeT takenT toT computeT isT increased,T andT thisT isT

whereT MachineT LearningT comesT intoT action,T toT helpT

peopleT withT largeT dataT inT minimumT time.T IfT bigT dataT

andT cloudT computingT areT gainingT importanceT forT theirT

contributions,T machineT learningT asT technologyT helpsT

analyzeT thoseT bigT chunksT ofT data,T easingT theT taskT ofT

dataT scientistsT inT anT automatedT processT andT gainingT

equalT importanceT andT recognition.T TheT techniquesT weT

useT forT dataT miningT haveT beenT aroundT forT manyT

years,T butT theyT wereT notT effectiveT asT theyT didT notT

haveT theT competitiveT powerT toT runT theT algorithms.T IfT

youT runT deepT learningT withT accessT toT betterT data,T theT

outputT weT getT willT leadT toT dramaticT breakthroughsT

whichT isT machineT learning.

II. TECHNIQUES OF MACHINE LEARNING The Machine Learning has three main Techniques those are as follow,

 Supervised learning  Unsupervised learning

 Semi-supervised and Reinforcement learning

i) Supervised Learning:

SupervisedT LearningT isT aT typeT ofT MachineT LearningT

usedT toT learnT modelsT fromT labeledT trainingT data.T ItT

allowsT usT toT predictT theT outputT forT futureT orT unseenT

data.

Figure 02 : Algorithm of Supervised Learning

SupervisedT LearningT Flow

 DataT Preparation  TrainingT Step

 EvaluationT orT TestT Step  ProductionT Deployment

TestingT theT Algorithm

GivenT belowT areT theT stepsT forT testingT theT algorithmT ofT

SupervisedT Learning.

1. OnceT theT algorithmT isT trained,T testT itT withT

testT dataT (aT setT ofT dataT instancesT thatT doT notT

appearT inT theT trainingT set).

2. AT well-trainedT algorithmT canT predictT wellT forT

newT testT data.

3. IfT theT learningT isT poor,T weT haveT anT underfitT

situation.T TheT algorithmT willT notT workT wellT

onT testT data.T RetrainingT mayT beT neededT toT

findT aT betterT fit.

4. IfT learningT onT trainingT dataT isT tooT intensive,T

itT mayT leadT toT overfittingT –T aT situationT whereT

theT algorithmT isT notT ableT toT handleT newT

testingT dataT thatT itT hasT notT seenT before.T TheT

techniqueT toT keepT dataT genericT isT calledT

regularization.

TypesT ofT SupervisedT Learning

GivenT belowT areT 2T typesT ofT SupervisedT Learning.

 Classification  Regression

ClassificationT SupervisedT Learning

LetT usT lookT atT theT classificationsT ofT SupervisedT

learning.

 AnswersT “WhatT class?”

 AppliedT whenT theT outputT hasT finiteT andT

International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-9S3, July 2019

RegressionT SupervisedT Learning

GivenT belowT areT someT elementsT ofT RegressionT

SupervisedT learning.

 AnswersT “HowT much?”

 AppliedT whenT theT outputT isT aT continuousT

number

 AT simpleT regressionT algorithm:T yT =T wxT +T b.T

Example:T theT relationshipT betweenT

environmentalT temperatureT (y)T andT humidityT

levelsT (x)

ii) UnsupervisedT Learning:

UnsupervisedT learningT isT aT termT usedT forT

HebbianT learning,T associatedT toT learningT withoutT aT

teacher,T alsoT knownT asT self-organizationandT aT methodT

ofT modellingT theT probabilityT densityT ofT inputs.

TheT unsupervisedT learningT toT createT clustersT

ofT heavenlyT bodies,T withT eachT clusterT containingT

objectsT ofT aT similarT nature.T UnsupervisedT LearningT isT

aT subsetT ofT MachineT LearningT usedT toT extractT

inferencesT fromT datasetsT thatT consistT ofT inputT dataT

withoutT labeledT responses.

ComparedT toT supervisedT learningT whereT

trainingT dataT isT labeledT withT theT appropriateT

classifications,T modelsT usingT unsupervisedT learningT

mustT learnT relationshipsT betweenT elementsT inT aT dataT

setT andT classifyT theT rawT dataT withoutT "help."T ThisT

huntT forT relationshipsT canT takeT manyT differentT

algorithmicT forms,T butT allT modelsT haveT theT sameT goalT

ofT mimickingT humanT logicT byT searchingT forT indirectT

hiddenT structures,T patternsT orT featuresT toT analyzeT newT

data.

TypesT ofT UnsupervisedT Learning

ThereT areT threeT typesT ofT UnsupervisedT LearningT are:

 Clustering

 VisualizationT Algorithms

 AnomalyT Detection

Clustering

TheT mostT commonT unsupervisedT learningT

methodT isT clusterT analysis.T ItT isT usedT toT findT

dataT clustersT soT thatT eachT clusterT hasT theT

mostT closelyT matchedT data.

VisualizationT Algorithms

VisualizationT algorithmsT areT unsupervisedT

learningT algorithmsT thatT acceptT unlabeledT dataT

andT displayT thisT dataT inT anT intuitiveT 2DT orT

3DT format.T TheT dataT isT separatedT intoT

somewhatT clearT clustersT toT aidT understanding.T

AnomalyT Detection

ThisT algorithmT detectsT anomaliesT inT dataT

withoutT anyT priorT training.T ItT canT detectT

suspiciousT creditT cardT transactionsT andT

differentiateT aT criminalT fromT aT setT ofT people.

iii) Semi-SupervisedT Learning

ItT isT aT hybridT approachT (combinationT ofT

SupervisedT andT UnsupervisedT Learning)T withT someT

labeledT andT someT non-labeledT data.

Semi-supervisedT learningT isT aT classT ofT

machineT learningT tasksT andT techniquesT thatT alsoT makeT

useT ofT unlabeledT dataT forT trainingT –T typicallyT aT smallT

amountT ofT labeledT dataT withT aT largeT amountT ofT

unlabeledT data.T Semi-supervisedT learningT fallsT betweenT

unsupervisedT learningT (withoutT anyT labeledT trainingT

data)T andT supervisedT learningT (withT completelyT

labeledT trainingT data).T ManyT machine-learningT

researchersT haveT foundT thatT unlabeledT data,T whenT

usedT inT conjunctionT withT aT smallT amountT ofT labeledT

data,T canT produceT considerableT improvementT inT

learningT accuracyT overT unsupervisedT learningT (whereT

noT dataT isT labeled),T butT withoutT theT timeT andT costsT

neededT forT supervisedT learningT (whereT allT dataT isT

labeled).TheT acquisitionT ofT labeledT dataT forT aT learningT

problemT oftenT requiresT aT skilledT humanT agentT (e.g.T toT

transcribeT anT audioT segment)T orT aT physicalT experimentT

(e.g.T determiningT theT 3DT structureT ofT aT proteinT orT

determiningT whetherT thereT isT oilT atT aT particularT

location).T TheT costT associatedT withT theT labelingT

processT thusT mayT renderT aT fullyT labeledT trainingT setT

infeasible,T whereasT acquisitionT ofT unlabeledT dataT isT

relativelyT inexpensive.T InT suchT situations,T

semi-supervisedT learningT canT beT ofT greatT practicalT value.T

Semi-supervisedT learningT isT alsoT ofT theoreticalT interestT

inT machineT learningT andT asT aT modelT forT humanT

learning.

III.T DATAT PREPROCESSINGT INT

MACHINET LEARNING

Here,T weT willT DiscussT DataT Preprocessing,T FeatureT

Scaling,T andT FeatureT EngineeringT inT detailT

Objectives

 RecognizeT theT importanceT ofT dataT preparationT

inT MachineT Learning

 IdentifyT theT meaningT andT aspectsT ofT featureT

engineering

 StandardizeT dataT featuresT withT featureT scaling  AnalyzeT datasetsT andT itsT examples

 ExplainT dimensionalityT reductionT withT

PrincipalT ComponentT AnalysisT (PCA)

DataT PreparationT inT

AT quickT briefT ofT DataT PreparationT inT MachineT

LearningT isT mentionedT below.

 MachineT LearningT dependsT largelyT onT testT

data.

 DataT preparationT isT aT crucialT stepT toT makeT itT

suitableT forT ML.

 AT largeT amountT ofT dataT isT generallyT requiredT

forT theT mostT commonT formsT ofT ML.

 DataT preparationT involvesT dataT selection,T

filtering,T transformation,T etc.

DataT PreparationT Process

TheT processT ofT preparingT dataT forT MachineT LearningT

algorithmTT comprisesTT theTT following:

 DataTT Selection  DataTT Preprocessing  DataT Transformation

DataT Selection

StepsT involvedT inT DataT SelectionT involves:

 ThereT isT aT vastT volume,T variety,T andT velocityT

ofT availableT dataT forT aT MachineT LearningT

problem.

 ThisT stepT involvesT selectingT onlyT aT subsetT ofT

theT availableT data.

 TheT selectedT sampleT mustT beT anT accurateT

representationT ofT theT entireT population.  SomeT dataT canT beT derivedT orT simulatedT fromT

theT availableT dataT ifT required.

 DataT notT relevantT toT theT problemT atT handT canT

beT excluded.

DataT Preprocessing

Let’sT understandT DataT PreprocessingT inT detailT below.

AfterT theT dataT hasT beenT selected,T itT needsT toT beT

preprocessedT usingT theT givenT steps:

1. FormattingT theT dataT toT makeT itT suitableT forT

MLT (structuredT format)

2. CleaningT theT dataT toT removeT incompleteT

variables

3. SamplingT theT dataT furtherT toT reduceT runningT

timesT forT algorithmsT andT memoryT

requirements.

DataT cleaningT atT thisT stageT involvesT filteringT itT basedT

onT theT followingT variables:

InsufficientT Data:T TheT amountT ofT dataT requiredT forT

MLT algorithmsT canT varyT fromT thousandsT toT millions,T

dependingT uponT theT complexityT ofT theT problemT andT

theT chosenT algorithm.

Non-RepresentativeT Data

TheT sampleT selectedT mustT beT anT exactT representationT

ofT theT entireT data,T asT non-representativeT dataT mightT

trainT anT algorithmT suchT thatT itT won'tT generalizeT wellT

onT newT testT data.

SubstandardT Data

Outliers,T errors,T andT noiseT canT beT eliminatedT toT getT aT

betterT fitmentT ofT theT model.T MissingT featuresT suchT asT

ageT forT 10%T ofT theT audienceT mayT beT ignoredT

completely,T orT anT averageT valueT canT beT assumedT forT

theT missingT component.

SelectingT theT rightT sizeT ofT theT sampleT isT aT keyT stepT inT

dataT preparation.T SamplesT thatT areT tooT largeT orT tooT

smallT mightT giveT skewedT results.

SamplingT Noise

SmallerT samplesT causeT samplingT noiseT sinceT theyT getT

trainedT onT non-representativeT data.T ForT example,T

checkingT voterT sentimentT fromT aT veryT smallT subsetT ofT

voters.

SamplingT Bias

LargerT samplesT workT wellT asT longT asT thereT isT noT

samplingT bias,T thatT is,T thenT theT rightT dataT isT picked.T

ForT example,T samplingT biasT wouldT occurT whenT

checkingT voterT sentimentT onlyT forT theT technicallyT

soundT subsetT ofT voters,T whileT ignoringT others.

DataT Transformation

TheT selectedT andT preprocessedT dataT isT transformedT

usingT oneT orT moreT ofT theT followingT methods:

1. Scaling:T ItT involvesT selectingT theT rightT featureT

scalingT forT theT selectedT andT preprocessedT data.

2. Aggregation:T ThisT isT theT lastT stepT toT collateT aT

bunchT ofT dataT featuresT intoT aT singleT one.

IV.T TYPEST OFT DATA

LabeledT DataT orT TrainingT Data

 ItT isT alsoT knownT asT markedT (withT values)T

data.

 ItT assistsT inT learningT andT formingT aT predictiveT

hypothesisT forT futureT data.T ItT isT usedT toT

arriveT atT aT formulaT toT predictT futureT behavior.  TypicallyT 80%T ofT availableT labeledT dataT isT

markedT forT training.

UnlabeledT Data

 DataT whichT isT notT markedT andT needsT

real-timeT unsupervisedT learningT isT categorizedT asT

International Journal of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, Volume-8 Issue-9S3, July 2019

TestT Data

 DataT providedT toT testT aT hypothesisT createdT

viaT priorT learningT isT knownT asT testT data.  TypicallyT 20%T ofT labeledT dataT isT reservedT forT

theT test.

ValidationT data

ItT isT aT datasetT usedT toT retestT theT hypothesisT (inT caseT

theT algorithmT gotT overT fittedT toT evenT theT testT dataT dueT

toT multipleT attemptsT atT testing).

TheT illustrationT givenT belowT depictsT howT totalT

availableT labeledT dataT mayT beT segregatedT intoT theT

trainingT dataset,T testT dataset,T andT validationT dataset.

V.T CONCLUSION:

Today,T itT isT inevitableT toT considerT andT useT

dataT miningT inT viewT ofT theT ever-increasingT amountT ofT

computerizedT businessT processesT andT theT hugeT amountT

ofT dataT toT beT analyzedT inT parallel.T ItT isT possibleT toT

makeT accurateT estimationsT orT predictionsT aboutT futureT

resultsT throughT applyingT machineT learningT techniquesT

toT theT dataT madeT availableT forT analysisT viaT dataT

mining.T ThisT studyT usedT inT variousT classificationT

techniquesT toT aT groupT ofT individualsT whoT areT inT theT

process.T SinceT usingT machineT learningT techniquesT inT

classificationT studiesT resultsT inT accurateT outcomesT

accompaniedT withT significantT savingT inT termsT ofT timeT

andT cost,T itT isT highlyT recommendedT toT makeT useT ofT

machineT learningT techniquesT inT DataT Processing.T ItT isT

consideredT thatT thisT studyT willT beT usefulT forT

organizationsT andT individualsT operatingT inT allT areasT ofT

workT thatT haveT goneT forT computerizationT ofT theirT

businessT processes.

REFERENCES:

1. J.T Qi,T P.T Yang,T G.T Min,T O.T Amft,T F.T Dong,T andT L.T

Xu,T "AdvancedT InternetT ofT ThingsT forT personalisedT

healthcareT systems:T AT survey,"T PervasiveT andT MobileT

Computing,T vol.T 41,T pp.T 132-149,T 2017.T

2. T.T Pawar,T N.T Anantakrishnan,T S.T Chaudhuri,T andT S.T P.T

Duttagupta,T "ImpactT ofT ambulationT inT wearable-ECG,"T

AnnalsT ofT BiomedicalT Engineering,T vol.T 36,T no.T 9,T pp.T

1547-1557,T 2008.T

3. L.-J.T KauT andT C.-S.T Chen,T "AT smartT phone-basedT

pocketT fallT accidentT detection,T positioning,T andT rescueT

system,"T IEEET journalT ofT biomedicalT andT healthT

informatics,T vol.T 19,T no.T 1,T pp.T 44-56,T 2015.T

4. M.T Ermes,T J.T Pärkkä,T J.T Mäntyjärvi,T andT I.T Korhonen,T

"DetectionT ofT dailyT activitiesT andT sportsT withT wearableT

sensorsT inT controlledT andT uncontrolledT conditions,"T IEEET

transactionsT onT informationT technologyT inT biomedicine,T

vol.T 12,T no.T 1,T pp.T 20-26,T 2008.T

5. D.T RiañoT etT al.,T "AnT ontology-basedT personalizationT ofT

healthcareT knowledgeT toT supportT clinicalT decisionsT forT

chronicallyT illT patients,"T JournalT ofT biomedicalT

informatics,T vol.T 45,T no.T 3,T pp.T 429-446,T 2012.T

6. M.T S.T Organization.T (2018,T NovT 2).T TopT 83T AIT

startupsT inT Healthcare.T Available:T

http://www.medicalstartups.org/top/ai.

7. Y.T GordienkoT etT al.,T "DeepT learningT withT lungT

segmentationT andT boneT shadowT exclusionT techniquesT forT

chestT x-rayT analysisT ofT lungT cancer,"T inT InternationalT

ConferenceT onT TheoryT andT ApplicationsT ofT FuzzyT

SystemsT andT SoftT Computing,T 2018,T pp.T 638-647:T

Springer.T

8. M.T T.T BrownT andT J.T K.T Bussell,T "MedicationT

adherence:T WHOT cares?,"T MayoT ClinicT proceedings,T vol.T

86,T no.T 4,T pp.T 304-314,T 2011.T

9. K.T Teng.T (2012,T May).T WhatT IsT PersonalizedT

Healthcare?T FromT patientsT toT medications,T oneT sizeT doesT

notT fitT all.T Available:T https://health.clevelandclinic.org/what-is-personalizedhealthcare.

10. A.T AraT andT A.T Ara,T "CaseT study:T IntegratingT IoT,T

streamingT analyticsT andT machineT learningT toT improveT

intelligentT diabetesT managementT system,"T inT 2017T

InternationalT ConferenceT onT Energy,T Communication,T DataT

AnalyticsT andT SoftT ComputingT (ICECDS),T 2017,T pp.T

3179-3182.T

11. R.T VargheeseT andT Y.T Viniotis,T "InfluencingT dataT

availabilityT inT IoTT enabledT cloudT basedT e-healthT inT aT

30T dayT readmissionT context,"T inT 10thT IEEET InternationalT

ConferenceT onT CollaborativeT Computing:T Networking,T

ApplicationsT andT Worksharing,T 2014,T pp.T 475-480.T

12. M.T D.T Naylor,T L.T H.T Aiken,T E.T T.T Kurtzman,T D.T M.T

Olds,T andT K.T B.T Hirschman,T "TheT importanceT ofT

transitionalT careT inT achievingT healthT reform,"T HealthT

affairs,T vol.T 30,T no.T 4,T pp.T 746-754,T 2011.T

13. S.T TilsonT andT G.T J.T Hoffman,T "AddressingT MedicareT

hospitalT readmissions,"T CongressionalT ResearchT Service,T

2012.T

14. S.T Shahrestani,T "AssistiveT IoT:T DeploymentT ScenariosT

andT Challenges,"T inT InternetT ofT ThingsT andT SmartT

Environments:T Springer,T 2017,T pp.T 75-95.T

15. M.T A.T HansonT etT al.,T "BodyT AreaT SensorT Networks:T

ChallengesT andT Opportunities,"T Computer,T vol.T 42,T no.T

1,T pp.T 58-65,T 2009.T

16. M.T J.T Rothman,T S.T I.T Rothman,T andT J.T BealsT IV,T

"DevelopmentT andT validationT ofT aT continuousT measureT

ofT patientT conditionT usingT theT ElectronicT MedicalT

Record,"T JournalT ofT biomedicalT informatics,T vol.T 46,T no.T