Nsryjdtyk

I'm dealing with multi-labelled data, and I would like to use stratify sampling. Let's assume I have 10 classes and let's call them 'ABCDEFGHIJ'. I have a dataframe with 10 columns corresponding to each of the label containing the rest of the info about the entries. I can extract those 10 columns in a n_entry*10 matrix that I will refer to as label_values

For instance, a line of label_values looks like [0,0,1,1,0,0,0,0,0,0] and this specific line means that the entry has Label C and Label D.

I would like to perform a split of my data in a training and validation set, and I would like to have the same proportion of each label in training and validation. To perform my splitting, I was using Sklearn train_test_split function (before my need to stratify), which happens to have an argument stratify. The current behaviour is to make the multi_label behaviour into a multiclass one (We consider [A,B] to be a brand new class totally different from class A and class B). As a result there are some classes with only 1 element, and this triggers an error :

ValueError("The least populated class in y has only 1"

                         " member, which is too few. The minimum"

                         " number of groups for any class cannot"

                         " be less than 2.")

coming from sklearn/model_selection/_split.py from the _iter_indices of the StratifiedShuffleSplit Class :

if np.min(class_counts) < 2:

        raise ValueError("The least populated class in y has only 1"

                         " member, which is too few. The minimum"

                         " number of groups for any class cannot"

                         " be less than 2.")

My fix was to override this method to delete this check. This works, and I get better repartition of my labels between train and validation. However, one of my labels with 2 elements is entirely in the train set. Is that normal?

Other question : Is this the good way to procede about this, or do you think there is a better way to get stratify train_test_split in the multi_label?