mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: SGD: Logistic regression package in Mahout
Date Tue, 30 Oct 2012 23:53:45 GMT
OK.  I am back up for air.

Rajesh,

As I am sure you know, most folks here contribute on their own time.  I
have been busy with my day job and unable to help with this until just now.

I just wrote a test case that looks at the Iris data set.  The results are
categorically different from yours.

That substantiates my original feeling that your encoding of the data is
problematic.  I will file a JIRA and attach a test case that you can look
at.  Then we can see what the differences are.


On Tue, Oct 23, 2012 at 1:28 AM, Rajesh Nikam <rajeshnikam@gmail.com> wrote:

> Hi,
>
> Is there development happening on fixing issue with SGD that generates
> models which are as good as random prediction?
>
> I am not sure why such issue is not noticed and raised by others ?
> May be this specific algo is not used in practical applications.
>
> Thanks,
> Rajesh
>
>
> >>
> >> On Tue, Oct 16, 2012 at 10:23 PM, Ted Dunning <ted.dunning@gmail.com
> >wrote:
> >>
> >>> Rajesh,
> >>>
> >>> In the testing that I did, I ran 100, 1000 and 10,000 passes through
> the
> >>> data.  All produced identical results.  Thus it isn't an issue of SGD
> >>> converging.
> >>>
> >>> I also did a parameter scan of lambda and saw no effect.
> >>>
> >>> I also did the standard thing in R with glm and got the expected
> >>> (correct)
> >>> results.
> >>>
> >>> I haven't looked yet in detail, but I really suspect that the reading
> of
> >>> the data is horked.  This is exactly how that behaves.
> >>>
> >>> On Tue, Oct 16, 2012 at 4:49 AM, Rajesh Nikam <rajeshnikam@gmail.com>
> >>> wrote:
> >>>
> >>> > Hi Ted,
> >>> >
> >>> > I was thinking, this might be due to having only 100 instances for
> >>> > training.
> >>> >
> >>> > So I have created test set with two classes having ~49K instances,
> >>> included
> >>> > all features as predictors.
> >>> > PFA sgd.grps.zip with test file.
> >>> >
> >>> > mahout trainlogistic --input /usr/local/mahout/trainme/sgd-grps.csv
> >>> > --output /usr/local/mahout/trainme/sgd-grps.model --target class
> >>> > --categories 2 --features 128 --types n --predictors a1 a2 a3 a4 a5
> a6
> >>> a7
> >>> > a8 a9 a10 a11 a12 a13 a14 a15 a16 a17 a18 a19 a20 a21 a22 a23 a24 a25
> >>> a26
> >>> > a27 a28 a29 a30 a31 a32 a33 a34 a35 a36 a37 a38 a39 a40 a41 a42 a43
> >>> a44 a45
> >>> > a46 a47 a48 a49 a50 a51 a52 a53 a54 a55 a56 a57 a58 a59 a60 a61 a62
> >>> a63 a64
> >>> > a65 a66 a67 a68 a69 a70 a71 a72 a73 a74 a75 a76 a77 a78 a79 a80 a81
> >>> a82 a83
> >>> > a84 a85 a86 a87 a88 a89 a90 a91 a92 a93 a94 a95 a96 a97 a98 a99 a100
> >>> a101
> >>> > a102 a103 a104 a105 a106 a107 a108 a109 a110 a111 a112 a113 a114 a115
> >>> a116
> >>> > a117 a118 a119 a120 a121 a122 a123 a124 a125 a126 a127
> >>> >
> >>> >
> >>> > mahout runlogistic --input /usr/local/mahout/trainme/sgd-grps.csv
> >>> --model
> >>> > /usr/local/mahout/trainme/sgd-grps.model --auc --confusion
> >>> >
> >>> > Still the results are similar, it classifies everything as class_1.
> >>> >
> >>> > AUC = 0.50
> >>> > confusion: [[*26563.0, 23006.0*], [0.0, 0.0]]
> >>> > entropy: [[-0.0, -0.0], [-46.1, -21.4]]
> >>> >
> >>> > I am not sure why this is failing all the time.
> >>> >
> >>> > Looking forward for your reply.
> >>> >
> >>> > Thanks
> >>> > Rajesh
> >>> >
> >>> >
> >>> >
> >>> > On Tue, Oct 16, 2012 at 3:57 AM, Ted Dunning <ted.dunning@gmail.com>
> >>> > wrote:
> >>> >
> >>> > > I would love to help and will before long.  Just can't do it in
the
> >>> first
> >>> > > part of this week.
> >>> > >
> >>> > > On Mon, Oct 15, 2012 at 6:28 AM, Rajesh Nikam <
> rajeshnikam@gmail.com
> >>> >
> >>> > > wrote:
> >>> > >
> >>> > > > Hello,
> >>> > > >
> >>> > > > I have asked below question on issue with using sgd on mahout
> >>> forum.
> >>> > > >
> >>> > > > Similar issue with sgd is reported by
> >>> > > >
> >>> > > >
> >>> > >
> >>> >
> >>>
> http://stackoverflow.com/questions/11221436/using-sgd-classifier-in-mahout
> >>> > > >
> >>> > > > Even below link has similar output:
> >>> > > >
> >>> > > > AUC = 0.57*confusion: [[27.0, 13.0], [0.0, 0.0]]*
> >>> > > > entropy: [[-0.4, -0.3], [-1.2, -0.7]]
> >>> > > >
> >>> > > >
> >>> > > >
> >>> >
> >>>
> http://sujitpal.blogspot.in/2012/09/learning-mahout-classification.html
> >>> > > >
> >>> > > > I am still wannder confusion how then this model works and
used
> by
> >>> > many ?
> >>> > > > Not able to get any points on how to use SGD that generates
> >>> effective
> >>> > > > model.
> >>> > > >
> >>> > > > Could someone point out what is missing in input file or
provided
> >>> > > > parameters.
> >>> > > >
> >>> > > > I appreciate your help.
> >>> > > >
> >>> > > > Below is description of steps that I followed.
> >>> > > >
> >>> > > > PF Attached uses input files for experiment.
> >>> > > >
> >>> > > > I am using Iris Plants Database from Michael Marshall. PFA
> >>> iris.arff.
> >>> > > > Converted this to csv file just by updating header:
> >>> iris-3-classes.csv
> >>> > > >
> >>> > > > mahout org.apache.mahout.classifier.
> >>> > > > sgd.TrainLogistic --input
> >>> > > /usr/local/mahout/trunk/*iris-3-classes.csv*--features 4 --output
> >>> > > /usr/local/mahout/trunk/
> >>> > > > *iris-3-classes.model* --target class *--categories 3*
> --predictors
> >>> > > > sepallength sepalwidth petallength petalwidth --types n
> >>> > > >
> >>> > > > >> it gave following error.
> >>> > > > Exception in thread "main" java.lang.IllegalArgumentException:
> Can
> >>> only
> >>> > > > call classifyScalar with two categories
> >>> > > >
> >>> > > > Now created csv with only 2 classes. PFA iris-2-classes.csv
> >>> > > >
> >>> > > > >> trained iris-2-classes.csv with sgd
> >>> > > >
> >>> > > > mahout org.apache.mahout.classifier.sgd.TrainLogistic --input
> >>> > > > /usr/local/mahout/trunk/*iris-2-classes.csv* --features 4
> --output
> >>> > > > /usr/local/mahout/trunk/*iris-2-classes.mode*l --target class
> >>> > > *--categories
> >>> > > > 2* --predictors sepallength sepalwidth petallength petalwidth
> >>> --types n
> >>> > > >
> >>> > > > mahout runlogistic --input
> >>> /usr/local/mahout/trunk/iris-2-classes.csv
> >>> > > > --model /usr/local/mahout/trunk/iris-2-classes.model --auc
> >>> --confusion
> >>> > > >
> >>> > > > AUC = 0.14
> >>> > > > confusion: [[50.0, 50.0], [0.0, 0.0]]
> >>> > > > entropy: [[-0.6, -0.3], [-0.8, -0.4]]
> >>> > > >
> >>> > > > >> AUC seems to poor. Now changed --predictors
> >>> > > >
> >>> > > > mahout org.apache.mahout.classifier.sgd.TrainLogistic --input
> >>> > > > /usr/local/mahout/trunk/*iris-2-classes.csv* --features 4
> --output
> >>> > > > /usr/local/mahout/trunk/*iris-2-classes.mode*l --target class
> >>> > > *--categories
> >>> > > > 2* --predictors sepalwidth petallength --types n
> >>> > > >
> >>> > > > mahout runlogistic --input
> >>> /usr/local/mahout/trunk/iris-2-classes.csv
> >>> > > > --model /usr/local/mahout/trunk/iris-2-classes.model --auc
> >>> --confusion
> >>> > > > --scores
> >>> > > >
> >>> > > > AUC = 0.80
> >>> > > > *confusion: [[50.0, 50.0], [0.0, 0.0]]*
> >>> > > > entropy: [[-0.7, -0.3], [-0.7, -0.4]]
> >>> > > >
> >>> > > > This model classifies everything as category 1 which of no
use.
> >>> > > >
> >>> > > > Thanks
> >>> > > > Rajesh
> >>> > > >
> >>> > > >
> >>> > > >
> >>> > > >
> >>> > >
> >>> >
> >>>
> >>
> >>
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message