A mistake about computing bias term

ajtulloch / svmpy

Basic soft-margin kernel SVM implementation in Python

252 stars 114 forks source link

A mistake about computing bias term #6

Open Jun-jie-Huang opened 5 years ago

Jun-jie-Huang commented 5 years ago

Hi, I've implemented a soft-margin SVM with kernel function and I referred to yours. I found there's a problem in your code to compute the bias term in the training process in this line.. Your code compute the $\sum_{i=1}^N alpha[i] * y[i] * K[i, j]$ by using the SVMPredictor which you set bias to 0. In this way, you compute the term AND use a sign function. But in the SVM algorithm, the sign function is not used in computing the bias term.

MatrixRanger98 commented 4 years ago

Yes. We shouldn't use sign function here. That is a major mistake rather than a bug. I highly suggest correcting it as there are already many webpages on svm quoting this repo.