Describe the bug
The problem appears on line 238 of file svm.py.In the original program, the formula for calculating the self.bias is as follows.
# Compute the bias
k = self.kernel(X_train, X_test)
SV_neg = y_train < 0
SV_pos = y_train > 0
self.bias = (-1 / 2) * (np.max(k[SV_neg[:, 0], :].T @ alpha[SV_neg]) + np.min(k[SV_pos[:, 0], :].T @ alpha[SV_pos]))
self.bias = y_train - np.sum(alpha * y_train * k, axis=1, keepdims=True)
self.bias = np.mean(self.bias)
The bias calculated in this way is incorrect and will cause errors in later predictions
Expected behavior
According to the formula I looked up, the correct calculation is as follows.
# Compute the bias
k = self.kernel(X_train, X_test)
SV_neg = y_train < 0
SV_pos = y_train > 0
kk=self.kernel(X_train, X_train)
self.bias = y_train - np.sum(alpha * y_train * kk, axis=1, keepdims=True)
self.bias = np.mean(self.bias)
Screenshots
Screenshot from the watermelon book "Machine learning" Zhou Zhihua section 6.2
Describe the bug
The problem appears on line 238 of file svm.py.In the original program, the formula for calculating the self.bias is as follows.
# Compute the bias
k = self.kernel(X_train, X_test)
SV_neg = y_train < 0
SV_pos = y_train > 0
self.bias = (-1 / 2) * (np.max(k[SV_neg[:, 0], :].T @ alpha[SV_neg]) + np.min(k[SV_pos[:, 0], :].T @ alpha[SV_pos]))
self.bias = y_train - np.sum(alpha * y_train * k, axis=1, keepdims=True)
self.bias = np.mean(self.bias)
The bias calculated in this way is incorrect and will cause errors in later predictions
Expected behavior
According to the formula I looked up, the correct calculation is as follows.
# Compute the bias
k = self.kernel(X_train, X_test)
SV_neg = y_train < 0
SV_pos = y_train > 0
kk=self.kernel(X_train, X_train)
self.bias = y_train - np.sum(alpha * y_train * kk, axis=1, keepdims=True)
self.bias = np.mean(self.bias)
Screenshots
Screenshot from the watermelon book "Machine learning" Zhou Zhihua section 6.2