Crammer and Singer (2001) have extended the binary SVM classifier to classification problems with more than two classes. The training problem of the Crammer-Singer multiclass SVM can be expressed as a QP
with variable where is the number of training examples and the number of classes. The kernel matrix is given by where is a kernel function and is the i’th row of the data matrix , and is an -vector with labels (i.e. ).
A custom solver for the multiclass support vector machine training
problem is available as a Python module
mcsvm. The module
implements the following function:
mcsvm(X, labels, gamma, kernel='linear', sigma=1.0, degree=1)¶
Solves the problem (1) using a custom KKT solver.
The input arguments are the data matrix (with the training examples as rows), the label vector , and the positive parameter .
Valid kernel functions are:
the linear kernel:
the polynomial kernel:
The kernel parameters and are specified using the input arguments sigma and degree, respectively.
Returns the function
classifier(). If is then classifier(Y) returns a list with as its k’th element
where is row of , is row of , and is the optimal solution of the QP (1).