Skip to content

Help understanding an error Line Search Fails

4 messages · Bill Poling, Ben Bolker

Original
Bill Poling
# 
Good afternoon. I hope I have provided enough info to get my question answered.
Using caret package I have been comparing models using my data, a training subset N=17357.

I have run PLS, RDA, GLM, and Boosted Logit based on a couple of tutorials.

http://dataaspirant.com/2017/01/19/support-vector-machine-classifier-implementation-r-caret-package/

https://cran.r-project.org/web/packages/caret/vignettes/caret.html

https://topepo.github.io/caret/model-training-and-tuning.html

However, when I get to trying svmLinear or svmRadial they both produce error: line search fails -1.614732 -0.257144 0.00001920624 0.00001369617 -0.00000001857456 -0.00000001542947 -0.000000000000568072

I have done some googling research but cannot find a definitive answer as to why this model does not work with my data but the other models do?

https://stackoverflow.com/questions/43267209/line-search-fails-when-training-a-model-using-caret

https://stackoverflow.com/questions/15895897/line-search-fails-in-training-ksvm-prob-model



Any advice would be appreciated.

Thank you

WHP

str(training)

# 'data.frame':17357 obs. of  7 variables:
#   $ SavingsReversed: num  0 0 0 0 0 ...
# $ productID      : num  3 3 3 3 3 1 3 3 3 1 ...
# $ ProviderID     : num  113676 114278 114278 114278 114278 ...
# $ ModCnt         : num  0 1 1 1 1 1 1 0 0 1 ...
# $ B2             : num  -1 -1 -1 -1 -1 -1 7 9 9 -1 ...
# $ B1a            : num  1 1 1 1 1 1 26 26 26 3 ...
# $ EditnumberI    : Factor w/ 2 levels "Bad","Good": 1 2 2 2 2 2 1 1 2 2 ...


head(training, n=25)

# SavingsReversed productID ProviderID ModCnt B2 B1a EditnumberI
# 1             0.00         3     113676      0 -1   1         Bad
# 5             0.00         3     114278      1 -1   1        Good
# 6             0.00         3     114278      1 -1   1        Good
# 7             0.00         3     114278      1 -1   1        Good
# 8             0.00         3     114278      1 -1   1        Good
# 10            0.00         1     114278      1 -1   1        Good
# 12          128.25         3     116641      1  7  26         Bad
# 13          159.60         3     116641      0  9  26         Bad
# 14            0.00         3     116641      0  9  26        Good
# 15            0.00         1     117280      1 -1   3        Good
# 16         1622.55         3     117439      1  9  26        Good
# 17           60.07         3     117439      1  9  26        Good
# 18            0.00         3     117439      0 -1   3        Good
# 19          190.00         3     117962      0  9  26        Good
# 20          372.66         3     119316      0  1  26         Bad
# 22            0.00         3     120431      1 -1   1        Good
# 25            0.00         3     121319      1  7  26         Bad
# 26           18.79         3     121319      1  7  26         Bad
# 27           23.00         3     121319      1  7  26         Bad
# 28           18.79         3     121319      1  7  26         Bad
# 29            0.00         3     121319      1  7  26         Bad
# 30           25.86         3     121319      2  7  26         Bad
# 31           14.00         3     121319      1  7  26         Bad
# 36          113.00         3     121545      1  1  26         Bad
# 37          197.20         3     121545      1  9  26         Bad


anyNA(training)
#[1] FALSE

My scripts

ctrl <- trainControl(
  method = "repeatedcv",
  repeats = 3,
  classProbs = TRUE,
  summaryFunction = twoClassSummary
)

set.seed(123)
svm_Linear <- train(EditnumberI ~., data = training,
                    method = "svmLinear",
                    trControl = ctrl,
                    preProcess = c("center", "scale"),
                    tuneLength = 10,
                    metric="ROC")
#warnings()
svm_Linear



set.seed(123)
svm_Radial <- train(EditnumberI ~., data = training,
                    method = "svmRadial",
                    trControl = ctrl,
                    preProcess = c("center", "scale"),
                    tuneLength = 10,
                    metric="ROC")
#warnings()
svm_Radial



line search fails -1.614732 -0.257144 0.00001920624 0.00001369617 -0.00000001857456 -0.00000001542947 -0.000000000000568072



WHP



Confidentiality Notice This message is sent from Zelis. ...{{dropped:13}}
Ben Bolker
# 
As far as I can tell none of the model types you're using fall under
the category of "mixed models" (linear/generalized linear models with
data  identified in known groups that are to be estimated by some form
of shrinkage estimator/"random effect").  (Please feel free to correct me!)

  By the way, I don't think it makes any sense to use "ProviderID" as a
*numeric* predictor variable ... that (and ProductID) are places where
you *might* actually want to use a mixed model.

  This looks like more of a CrossValidated question - note that you'll
have to provide a *reproducible* example in order to get help ...

  cheers
    Ben Bolker
On 2018-12-05 12:45 p.m., Bill Poling wrote:
Bill Poling
# 
Good morning Ben and thank you for your response.

Yes, had not considered this a sig-mixed-models question but was unsure of where to start my questions.

How would I make available a reproducible example for further help?

If you suggest that this may be more suitable data for mixed model I would like to pursue that.

Appreciate your help Sir, thank you.

WHP




From: R-sig-mixed-models <r-sig-mixed-models-bounces at r-project.org> On Behalf Of Ben Bolker
Sent: Wednesday, December 5, 2018 12:48 PM
To: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] Help understanding an error Line Search Fails


As far as I can tell none of the model types you're using fall under
the category of "mixed models" (linear/generalized linear models with
data identified in known groups that are to be estimated by some form
of shrinkage estimator/"random effect"). (Please feel free to correct me!)

By the way, I don't think it makes any sense to use "ProviderID" as a
*numeric* predictor variable ... that (and ProductID) are places where
you *might* actually want to use a mixed model.

This looks like more of a CrossValidated question - note that you'll
have to provide a *reproducible* example in order to get help ...

cheers
Ben Bolker
On 2018-12-05 12:45 p.m., Bill Poling wrote:
_______________________________________________
mailto:R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

Confidentiality Notice This message is sent from Zelis. ...{{dropped:13}}
Ben Bolker
# 
You could ask your question on r-help or StackOverflow or
CrossValidated (https://stats.stackexchange.com). r-help is mostly for R
questions (obviously).  StackOverflow might be best, as this is
primarily a programming question (and you're already looking on SO for
answers ...)

 There's a whole lot of information & advice on constructing
reproducible examples here:

https://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example

  If you choose to post on SO or CV, make sure to read the info on "how
to post"; for R-help, read https://www.r-project.org/posting-guide.html

  good luck,
     Ben Bolker
On 2018-12-06 7:36 a.m., Bill Poling wrote: