I got asked as an assignment to develop a custom polynomial (degree = 3,4,5) kernel for SVM and compare its accuracy to the in-built poly kernel of sklearn (should be almost the same). I tried to follow the polynomial kernel definition but my results seems not to be quite similar, this is my code: def poly_kernel_fn(X, Y): # Implement a polynomial kernel # - args: 2 numpy arrays of shape [n_samples, n_features] # - returns: computed kernel matrix of shape [n_samples, n_samples] K = np.zeros((X.shape[0],Y.shape[0])) K = (X.dot(Y.T) + 1)**4 return K clfpoly = svm.SVC(kernel='poly', degree=4) clfpoly.fit(X_train, y_train) zpoly = clfpoly.predict(X_test) print("The accuracy with in-built 3D polynomial kernel is: ",accuracy_score(y_test, zpoly)*100,"%") clf = svm.SVC(kernel=poly_kernel_fn) clf.fit(X_train, y_train) z = clf.predict(X_test) print("The accuracy with custom rbf kernel is: ",accuracy_score(y_test, z)*100,"%") The accuracy results is the following: The accuracy with in-built 4D polynomial kernel is: 56.99999999999999 % The accuracy with kernel is: 59.0 % If I change the polynomial grade to 3 or 5 it changes even more, so I do not know if I am doing something wrong or simply is not possible to match the in-built accuracy. Continue reading...
