Search code examples
parameterssvmmargins

Negative sign of the bias of a SVM model

I have a doubt regarding the relationship between bias and the parameter C in a SVM (C=inverse of regularization parameter λ). I am training a model in MATLAB and I need to set C. I know this rule: large C brings lower bias and higher variance, while small C brings higher bias and lower variance.

First, this rule doesn't seem to be true in my model, when I set a C bigger than 50. The bias returned is always negative for every C I choose. The problem is that with C>50 I obtain higher biases (in absolute value) than the ones with, for example, very low C (<1).

Secondly, the bias should be considered and reported as its absolute value or with its sign?

Thank you for your help,

MP

Solution

  • After posting the question, of course, I've solved my doubt!

    If someone is interested, the issue is just a word misinterpretation. In that context, with the word bias MATLAB does not mean an estimate of the performance of the model, but the distance hyperplane-features space origin. The documentation is not clear enough for my poor knowledge!

    Bye