I'm using Libsvm to train the model but the result is always not so good. now I'm reading this lecture's document:


the lecture mentioned "Three-way data splits", when we train a model, we should have one "training data" and two "testing data", the training data form the model, the first testing data adjust the parameter, then the second testing data get the accurancy.

it's reasonable for me and I want to find more information about the method (and if there is a source code of this method. it is difficult to me to modify the grid.py of libsvm to this Three-way data splits)

however, all the information I can search in the internet is only the lecture's note. I'm wondering there is other name of this method.

thanks for any information!

  • $\begingroup$ Couldn't you find out any other reading on the training/validation/test dataset splitting on the internet? $\endgroup$ – tagoma Sep 19 '17 at 6:48
  • $\begingroup$ ok, so the word "validation set" is what I missing. thanks very much! $\endgroup$ – ggininder Sep 19 '17 at 7:20

I usually hear it being referred to as "splitting your data up into, a training, test, and validation set".

Google seems to be quite receptive to the phrase, "splitting data machine learning validation" which suggest that a common term for it is just splitting the data, here into three categories, as you suggested. There is, at least, a quite wide range of sources using this term.

A related technique, with an actual name, is k-fold cross validation where the test and training set are initially combined, but later split up into k folds where one of them functions as the test set. After testing this, the test set is shifted to the next fold repeat k times. When you have k test scores they are usually averaged to get the final test score. Here it's less important to have a validation set as it's harder for the model to overfit.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.