[ENH] Add a label encoder option when loading dataset

[baraline]

Describe the feature or idea you want to propose

When loading datasets, we often get class labels as strings, which is not the best case when wanting to do plotting or other operation (e.g. plotting samples coloured by their class label).

It would be nice to have an option as a boolean parameter to apply (and return) a label encoder directly when loading, instead of having to do it outside this function every time we want too.

Describe your proposed solution

Add a boolean parameter (default to false to not change default behaviour) that applies and returns a label encoder to the class labels when loading classification datasets.

Some thinking might be needed on how to handle properly the case where we load the training data, and then need to re-use the label encoder for the testing data.

Describe alternatives you've considered, if relevant

No response

Additional context

I often need to do this when loading classification datasets, so I might not be the only one (hopefuly!)

[TonyBagnall]

good idea, so would that be in just load_classification or also load_to_ts_file? Do you need to return the label encoder?

[baraline]

I guess we can apply it to any function returning discrete targets ?

For the encoder, that's the tricky part, for example, returning the label encode when loading train, means we need to supply it when loading tests, something like

X_train, y_train, le = load_classification(dataset_name, split="train", use_label_encoder=True)
X_test, y_test, _ = load_classification(dataset_name, split="test", use_label_encoder=le)

Just feels weird and convoluted to me.

I think the simpler solution would be to have something like :

X_train, X_test, y_test, y_train = load_classification_train_test(dataset_name, use_label_encoder=True)

Which would internally handle the label encoding. It would seem like defining a new function would be best, I don't think adding more parameter combinations to the existing one would make thing easier.

[MatthewMiddlehurst]

Did not consider that before giving my thumbs up 🙂. New function would be best IMO if this is wanted, would not agree to the changes in the first code block.