資源描述：
《4 - Supervised Learning - Bayesian Classification》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、MachineLearning,MachineLearning(extended)4–SupervisedLearning:BayesianClassificationKashifRajpootk.m.rajpoot@cs.bham.ac.ukSchoolofComputerScienceUniversityofBirminghamOutline?Supervisedlearning?Classification?Probabilisticvsnon-probabilistic?Generativevsd
2、iscriminative?Refresher:probability?Bayesianclassification?Na?veBayesclassification2?GaussianclassificationSupervisedlearning?Regression?Minimisedloss(e.g.leastsquares)?Maximumlikelihood?Classification?Generative(e.g.Bayesian)?Instance-based(e.g.k-NN)?Dis
3、criminative(e.g.SVM)3Classification?AsetofNobjectswithattributes(usuallyvector)???Eachobjecthasanassociatedtargetlabel???Binaryclassification??∈0,1or??∈?1,1?Multi-classclassification??∈1,2,…,??Classifierlearnsfrom?1,?2,…,??and?1,?2,…,??sothatitcanlatercla
4、ssify????4Probabilisticvsnon-probabilisticclassification?Probabilisticclassifiersproduceaprobabilityofclassmembership?Non-probabilisticclassifiersproduceahardassignment5Probabilisticvsnon-probabilisticclassification?Probabilitiesprovideuswithmoreinformati
5、on??????=1=0.6ismoreusefulthan????=1?Confidencelevel?Particularlyimportantwherecostofmisclassificationishighandimbalanced?Diagnosis:tellingadiseasedpersontheyarehealthyismuchworsethantellingahealthypersontheyarediseased6Generativevsdiscriminativeclassific
6、ation?Generativeclassifiersgenerateamodelforeachclass,basedontrainingsamplesavailable?Dataineachclasscanbeseenasgeneratedbysomemodel?Fornewtestsamples,theyassignthesesamplestotheclassthatsuitsbest(e.g.byprobabilitymeasure)?Incontrast,discriminativeclassif
7、iersattempttoexplicitlydefinethedecisionboundarythatseparatestheclasses?Intuitively,thesemethodsareforbinaryclassproblemsbutcanbeextendedtomulti-classproblems7Bayesianclassifier?AclassifierbuiltonBayesrule?Buildsaprobabilisticmodelofthedata,embeddingprior
8、knowledge?Allowsustoextractpriorknowledgefromobserveddata?Generativeapproach?Buildsamodelfromtrainingobjects?Anynewobjectscanbeclassifiedbasedontheprobabilisticmodelspecification8Refresher:probability?Conditionalpro
