SVM C Parameter. higher C values allow for 'softer' boundaries

balaji_k08

I am taking rapid miner analyst certification in a couple of days.

 

In Rapid Miner documentation it says, higher C values allow for 'softer' boundaries and lower values create 'harder' boundaries. Is this correct ?  Ideally large C values should lead to hard margins right ? Pls Suggest

IngoRM

Hi,

 

I guess this depends a bit on how you define the terms "hard" and "soft" here :-)

 

Think about a RBF kernel which can be non-linear.  "Soft" in the documentation means that you can "bend" or "move" the hyperplace more, i.e. you can bring it into a shape which reduces the training error.  Which is exactly what happens for higher values of C.  While a "hard" boundary tends to stay in its shape and cannot be moved that easily.  This can therefore increase the training error.

 

I would probably somewhat avoid to use those fuzzy terms altogether now but I am pretty sure that is what is meant by the documentation.

 

Hope this helps,

Ingo