C <= 0! While Using Support Vector Machine for Classification

Question

I am working on lead scoring classification problem, wherein the model i am building would help highlight the customers likely to purchase based on their online activity. Below is the brief snapshot of the data i have:

'data.frame':   889 obs. of  20 variables:
 $ daysInSystem        : num  142 86 102 1 1 153 131 1 1 1 ...
 $ sessionCount        : num  25 1 6 1 1 1 12 1 1 2 ...
 $ maxSessionDuration  : num  2687 105 258 223 1821 ...
 $ totalSessionDuration: num  19059 105 319 223 1821 ...
 $ usedCarPageCount    : num  2 0 0 4 3 2 5 0 0 0 ...
 $ Model.Page.count    : num  155 2 22 1 5 0 40 15 5 7 ...
 $ Variant.Page.count  : num  31 0 0 1 0 0 4 0 0 7 ...
 $ OEM.Page.count      : num  0 0 0 0 0 0 0 0 0 0 ...
 $ OEMCount            : num  11 1 2 1 1 0 3 2 1 1 ...
 $ modelCount          : num  33 1 3 1 1 0 7 10 1 1 ...
 $ carVariantLeadCount : num  45 1 0 1 1 0 1 3 0 0 ...
 $ ORP_LC              : num  49 1 3 4 7 2 10 7 2 1 ...
 $ UPCOMING_ALERT_LC   : num  0 0 0 0 0 0 0 0 0 0 ...
 $ BROCHURE_LC         : num  0 0 0 0 0 0 0 0 0 0 ...
 $ DCB_LC              : num  0 0 0 0 1 0 0 0 0 0 ...
 $ OFFER_LC            : num  0 0 0 0 0 0 0 0 0 0 ...
 $ RECOMENDATION_LC    : num  0 0 0 0 0 0 0 0 0 0 ...
 $ distinctModels      : num  19 3 2 3 4 1 7 5 1 1 ...
 $ distinctBrands      : num  9 2 2 3 3 1 3 5 1 1 ...
 $ TR                  : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ...

"TR" is my dependent or class variable. I am trying to use SVM to create a classification model using e1071 package. Following are codes i am using:

> svm_train <- svm(TR ~ ., data = train_randomsmote, kernel = 'linear', cost = 1)
> svm_train

Call:
svm(formula = TR ~ ., data = train_randomsmote, kernel = "linear", cost = 1)


Parameters:
   SVM-Type:  C-classification 
 SVM-Kernel:  linear 
       cost:  1 
      gamma:  0.05263158 

Number of Support Vectors:  621

> tune.out <- tune(svm, TR~ ., data = train_randomsmote, kernel = 'linear', ranges = list(cost = c(.001,0.01,0,1,5,10,100)))
Error in svm.default(x, y, scale = scale, ..., na.action = na.action) : 
  C <= 0!

My Question is what is the meaning of the error - Error in svm.default(x, y, scale = scale, ..., na.action = na.action) : C <= 0! - and is there any way to solve this. I know C pertains to cost, and is used to indicate the distance between between two classes. But i am not able to understand the error.

Let me know if more details is required. I cannot share the data.

Answer 1

The cost parameter ($C$) in an SVM places an upper bound on the norm of the weights. It is a regularisation parameter, where lower means more regularised and greater means less regularised.

Low $C$ means the norm of the weights has to be very small, leading to a simpler model. High $C$ is the opposite. $C=0$ would imply that the norm of the weights must be zero and that there is no model at all, and so svm throws an error. You have a 0 in your gridsearch.

tune.out <- tune(svm, TR~ ., 
                 data = train_randomsmote, 
                 kernel = 'linear', 
                 ranges = list(cost = c(.001,0.01,0,1,5,10,100)))
# Right here -------------------------------------^