在Sci-kit Learn中将参数解析为SVM的自定义内核功能

Question

I followed the tutorial SVM with custom kernel and tried to use custom kernel in SVM. For example, I implement the polynomial kernel function as follows:

   def poly_kernel(x, y):
       degree = 3
       return np.dot(x, y.T) ** 3

Then the result seem very similar to the original 'poly' with degree of 3. However it comes with a problem that I do not know how to parse the degree as a parameter to the kernel function.

For example, I build the Support Vector Regression as follows:

    # X is some data
    # y is some target
    svr = SVR(kernel=poly_kernel, C=1e3, degree=4)
    y = svr.fit(X, y).predict(X)

It does not seem to parse the parameter to the kernel correctly. I also tried named arguments in the kernel function

    def poly_kernel(x, y, **kwargs):
        degree = 3
        try:
            degree = kwargs.get('degree')
        except:
            pass
        return np.dot(x, y.T) ** 3

But it does not work.

So is there any way to parse the parameters correctly in this case?

Thanks in advance.

Answer 1

The kernel function can be dynamically constructed in this case. We can use lambda get a anonymous function as a variable.

For example:

    def linear_kernel(c = 0):
        return lambda x, y: np.dot(x, y.T) + c

When we want to use it, we just do:

    lkf = linear_kernel(c=20)
    svr_linear = SVR(kernel=lkf)
    y_linear = svr_linear.fit(X, y).predict(X)

and I just ignore the parameters when I call SVR().

However I am not sure if it is a dirty way, but it works at least.