How can one decide on using a linear or non linear classifier for the dataset in SAS Miner?

[sdhilip]

Hi,

How can one identify whether the samples are linearly separable or not before applying a binary classifier? (or) when should I opt for linear SVM and non-linear SVM? 

In one of my works, I applied linear SVM and kernel SVM (Polynomial) for the same dataset in SAS Miner. Kernel SVM performs better in terms of accuracy. Does it mean that Dataset is not linearly separable?

Is there any way we can check the datasets visually before applying SVM? My dataset size around 45000 (70: training & 30: validation)

Regards,

Dhilip

[koyelghosh]

Non-linear SVM classifier is supposed to give better accuracy over linear. Is your linear SVM far worse than the non-linear SVM/Kernel? May be delta accuracy is not large enough to discard the linear classifier. 

There is a nice discussion here (https://stats.stackexchange.com/questions/73032/linear-kernel-and-non-linear-kernel-for-support-vect...)

[koyelghosh · Accepted solution]

As far as your main question "Is there any way we can check the datasets visually before applying SVM?" is concerned, I am not sure if you can analyze your full data with all the features (if the number of features are more than 3). Simply because it become harder and harder to visualize once the dimensions go beyond 3.
I am not sure if I answered your question. I am sorry if I did not.
Best wishes.