[pdp] Extend modeling functionality

[bdewilde]

changes

adds functions (and unit tests) for common modeling-adjacent tasks: splitting datasets into exclusive subsets, and computing sample weights to deal with class imbalance

context

I've had to do this stuff "custom" for more than one school, but it's the same logic, so should be in the shared lib code.

Follows PR #40

questions

Are there any other tasks like this to include?

[vishpillai123]

Sample weight function looks to be an extension of sklearn.utils.class_weight.compute_sample_weight, so looks great to me.

For the compute_data_splits() function, I was curious why using a random generator to create the dataset splits was necessary, instead of just using sklearn.model_selection.train_test_split? I have seen train_test_split commonly utilized across academia/industry, so just curious of your choice of implementing the rng library here. Not really requesting changes and I'm definitely open to your thoughts here, just thought to ask. :)

[bdewilde]

As the name suggests, train_test_split() only splits the dataset into 2 components -- "train" and "test" -- while we want 3 -- the additional "validate" split. It's possible to apply train_test_split() twice, but that's slower and clunky. Using the underlying numpy functionality directly is faster and accomplishes the same thing. :)

[vishpillai123]

As the name suggests, train_test_split() only splits the dataset into 2 components -- "train" and "test" -- while we want 3 -- the additional "validate" split. It's possible to apply train_test_split() twice, but that's slower and clunky. Using the underlying numpy functionality directly is faster and accomplishes the same thing. :)

Ah, I thought that you could create a validation split and I could have sworn I've used it before, but I'm probably mixing up something else I've used. Ok, makes sense!