Can I split train and test data manually?

[BlockChanZJ]

can I use gk.fit_transform(graphs)，（train data and test data are included） then use the sub-matrix of train and test data to run SVM？

[giannisnik]

Hi @BlockChanZJ , Yes, you can do that. But that way, you will also compute the kernel values between the test samples (the SVM classifier doesn't in fact need those kernel values). Thus, if the number of test samples is very large, it will be computationally more expensive than using first fit_transform() and then transform().

[BlockChanZJ]

@giannisnik Thanks for your answer! I mean fit_transform is only for calculate kernel matrix? If so, I don't need to run multi times fit_transform when using k-fold cross validation.

[giannisnik]

Yes, you are right. For k-fold cross validation, you can compute the kernel matrix once (using fit_transform on all graphs) and then at each fold extract the corresponding training and test matrices.

[BlockChanZJ]

Thanks for your answer! It helps me a lot! 👍