Hi all. Is it possible to "convince" matlab fitcsvm function to use a well-defined (not random) subset of the sample vectors for training (leaving the others for testing)? Not simply a random percentage, as set by the "'Holdout', value" pair, but a list of indices (decided by me) to exactly choose the desired samples from the whole dataset.
If I could have a percentage equal to 0 in Holdout, it would do, because the machine would be trained on all the input vectors, then I'd use "predict" on the test subset.
This is absolutely necessary for my code, because I must be able to use the same sample subsets for training etc of different classifiers.
To be clearer, when using a neural network (by patternnet, in the Neural Network Toolbox), I can decide which sample vectors to use for training, validation, and test, by net.divideFcn = 'divideind', then setting manually the indices to be used for training etc.
Thanks.
Best regards.
Giorgio
Replying to myself... I was not understanding but perhaps now it is clear.
By training fitcsvm with a simple fitcsvm(x,y) I can train the machine with the whole set of data (everything is used as the training set). The trained machine can then be applied to a new (test) data set by the predict function. This is exactly what I need.
Only if I set an option such as 'CrossVal', 'CVPartition', etc, I get a ClassificationPartitionedModel, with a number of machines trained accordingly. Otherwise, I get a ClassificationSVM classifier.
It was simple...
Bye
Giorgio
I was about to propose you to split the data into two sets. your reply shows the answer,
best regards
Hi Nelly, thank you very much!
Yes, I found the solution after a hard (and stupid) struggle :-D... it was easy and direct....
My svm's are now happily working, with AUC above 0.96 after all :-).
It is a problem of Brain-Computer Interfacing....
Thanks again
Best regards
Giorgio
Hi Ariadne,
thank you very much for your kind reply!
My problem was how to set fitcsvm so as to understand that it should partition data according to my own sample choice (so, not simply according to train/validation percentages, but to actual previously-defined sample groups).
I found that if no validation approach is suggested to fitcsvm (cross validation, hold out etc), the machine is trained on the whole data set. So, as you rightfully suggest, I split data before calling fitcsvm, I then passed only the training data, and later I used "predict" on the left-out samples.
Thank you again.
Best regards
Giorgio
- Badges
- Science topic