In case before i was on purpose try to get a bad model just to check if
validation is wrong/bugged.In real case when i choose predictors variables
correct validation is even weirder.
cmd to train adaptive log.reg.(real deal):
mahout trainAdaptiveLogistic --input DataFraud100kTrening.csv --output
model --target fraudRisk --predictors balance numTrans numIntlTrans
creditLine --types text --passes 50 --categories 2
validation cmd: mahout validateAdaptiveLogistic --input
DataFraud100kTest.csv --model model --auc --confusion
and to output is:
Log-likelihood:Min=-5.61, Max=-0.00, Mean=-0.16, Median=-0.01
AUC = 0.12
=======================================================
Confusion Matrix
-------------------------------------------------------
a b <--Classified as
18812 0 | 18812 a = 0
0 1187 | 1187 b = 1
Entropy Matrix: [[-4.3, -0.6], [-0.0, -0.1]]
using mahout 0.9,data has 100k point
variables:"custID","gender","state","cardholder","balance","numTrans","numIntlTrans","creditLine","fraudRisk".
On standard log.reg. everything is fine AUC is around ~0.75.Sorry for my
bad english :).
On Fri, Jul 11, 2014 at 5:53 AM, Ted Dunning wrote:
> THis is confusing for sure. The AUC says that you have no predictive
> power, but the confusion matrix says you have a perfect solution.
>
> Can you say more about what you did and what data you used?
>
>
>
>
> On Thu, Jul 10, 2014 at 6:14 AM, fqsbs1 . wrote:
>
> > Hi,i have one question.Is validation of adaptive log.reg. bugged becouse
> i
> > get this results?
> >
> > Log-likelihood:Min=-0.69, Max=-0.69, Mean=-0.69, Median=-0.69
> >
> > AUC = 0.48
> >
> > =======================================================
> > Confusion Matrix
> > -------------------------------------------------------
> > a b <--Classified as
> > 18812 0 | 18812 a = 0
> > 0 1187 | 1187 b = 1
> >
> >
> >
> > Entropy Matrix: [[-0.7, -0.2], [-0.7, -0.2]]
> >
>