0

When I run an SVM with ksvm from the kernlab package, all the outputs from the predict command on my final model are scaled. I know this is because I initiate scaled = T but I also know scaling your data is preferred in SVM modeling. How can I easily tell ksvm to return non-scaled predictions? If not, is there a way to just manipulate the predicted scaled values to raw values? Thank you, code is below:

svm1 <- ksvm(Y ~ 1
            + X1
            + X2
            , data = data_nn
            , scaled=T
            , type = "eps-svr"
            , kernel="anovadot"
            , epsilon = svm1_CV2$bestTune$epsilon
            , C = svm1_CV2$bestTune$C
            , kpar = list(sigma = svm1_CV2$bestTune$sigma
                          , degree=  svm1_CV2$bestTune$degree)  
            ) 

#Analyze Results
data_nn$svm_pred <- predict(svm1)
LyzandeR
  • 37,047
  • 12
  • 77
  • 87
gtnbz2nyt
  • 1,465
  • 3
  • 17
  • 33

1 Answers1

2

From the documentation:

argument scaled:
A logical vector indicating the variables to be scaled. If scaled is of length 1,
the value is recycled as many times as needed and all non-binary variables are scaled. 
Per default, data are scaled internally (both x and y variables) to zero mean and 
unit variance. The center and scale values are returned and used for later predictions.

SOLUTION NO.1

Let's see the following example:

#make random data set
y <- runif(100,100,1000) #the response variable takes values between 100 and 1000
x1 <- runif(100,100,500)
x2 <- runif(100,100,500)
df <- data.frame(y,x1,x2)

Typing this:

svm1 <- ksvm( y~1+x2+x2,data=df,scaled=T,type='eps-svr',kernel='anovadot')

> predict(svm1)
               [,1]
  [1,]  0.290848927
  [2,] -0.206473246
  [3,] -0.076651875
  [4,]  0.088779924
  [5,]  0.036257375
  [6,]  0.206106048
  [7,] -0.189082081
  [8,]  0.245768175
  [9,]  0.206742751
 [10,] -0.238471569
 [11,]  0.349902743
 [12,] -0.199938921

Makes scaled predictions.

But if you change it to the following according to the documentation from above:

svm1 <- ksvm( y~1+x2+x2,data=df,scaled=c(F,T,T,T),type='eps-svr',kernel='anovadot')
#I am using a logical vector here so predictions will be raw data.
#only the intercept x1 and x2 will be scaled using the above.
#btw scaling the intercept (number 1 in the formula), actually eliminates the intercept.

> predict(svm1)
           [,1]
  [1,] 601.2630
  [2,] 599.7238
  [3,] 599.7287
  [4,] 599.9112
  [5,] 601.6950
  [6,] 599.8382
  [7,] 599.8623
  [8,] 599.7287
  [9,] 601.8496
 [10,] 599.0759
 [11,] 601.7348
 [12,] 601.7249

As you can see this is raw data predictions.

SOLUTION NO.2

If you want to scale the y variable in the model you ll need to unscale the predictions yourself.

Before the model:

Calculate the mean and std before running the model:

y2 <- scale(y) 
y_mean <- attributes(y2)$'scaled:center' #the mean
y_std <- attributes(y2)$'scaled:scale'   #the standard deviation

Convert the predictions to raw:

svm1 <- ksvm( y~1+x2+x2,data=df,scaled=T,type='eps-svr',kernel='anovadot')

> predict(svm1) * y_std + y_mean
           [,1]
  [1,] 654.3604
  [2,] 522.3578
  [3,] 556.8159
  [4,] 600.7259
  [5,] 586.7850
  [6,] 631.8674
  [7,] 526.9739
  [8,] 642.3948
  [9,] 632.0364
 [10,] 513.8646
 [11,] 670.0349
 [12,] 524.0922
 [13,] 673.7202

And you got raw predictions!

LyzandeR
  • 37,047
  • 12
  • 77
  • 87