基本sydfn
precision-recallcurvessummarizethetrade-offbetweenthetruepositiverateandthepositivepredictivevalueforapredictivemodelusing
precision-recall curves 3360 imbalanced数据集。
预处理可移植性
In a classification problem,wemaydecidetopredicttheclassvaluesdirectly.alternately, itcanbemoreflexibletopredicttheprobability itcanprovidethecapabilitytochooseandevencalibratethethethreshowtointerpretheppabilitytthephetophetytothos
twotypesoferrorswhenmakingapredictionforabinary/two-classclassificationproblem 3360
fn : predictnoeventwhentherewasanevent
acommonwaytocomparemodelsthatpredictprobabilitiesfortwoclassesistousearoccurve。
sensitvity : truepositiverate=TP/(yqdbdfn ) )。
假定位速率=Fp/(舒适画笔)=1-特定
规格=TN /舒适的画笔
accuracy=(TPTN )/(yqdbdTN FP FN ) )。
recall=TP/(TP FP )
presicion and recall are trade off。
if we want to cover more sample,then it ' seasiertomakemistakes-high recall-low precision
ifwehaveconcernedmodel-low recall-high precision
smallervaluesonthex-axisoftheplotindicatelowerfalsepositivesandhighertruenegatives。
largervaluesonthey-axisoftheplotindicatehighertruepositivesandlowerfalsenegatives
when we predict a binary outcome,itiseitheracorrectprediction (true position ) or not (false positive ).thereisatensionbetweeenthen
askilfulmodelwillassignahigherprobabilitytoarandomlychosenrealpositiveoccurrencethananegativeoccurrenceonaverage.thisiswhatwhatwer lhasskill.generally,skilfulmodelsarerepresentedbycurvesthatbowuptothetopleftoftheplot。
ano-skillclassifierisonethatcannotdiscriminatebetweentheclassesandwouldpredictarandomclassoraconstantclassinallcases.amode epoint (0.5,0.5 ).amodelwithnoskillateachthresholdisrepresentedbyadiagonallinefromthebottomleft
amodelwithperfectskillisrepresentedatapoint (0,1 ).amodelwithperfectskillisrepresentedbyalinethattravelsfromthebotttomleftomlefteftefttttom
anoperatormayplottheroccurveforthefinalmodelandchooseathresholdthatgivesadesirablebalancebetweeenthefalsepositivesandfalsenegatategatatation
控制召回和修复。
recall-risk-sensitivity-truepositiverate的希望是1