Splitset.make(max_imbalance:float)

[aiqc]

Background

Splitset.make is where the sample indices that make up the splits are defined.

Problem

When labels are not balanced, the network gets biases toward the majority classes and performs poorly on minority classes.

Need a way to downsample majority classes in order to balance categorical labels prior to split creation

Solution

• New argument max_imbalance:float (greater than 1.0) for Splitset.make
• Reference the Balance section in this gist: https://gist.github.com/aiqc/d8d4b5e74a8811b3d8657c65cb3c6e7f
• Programmatically figure out which class is the minority
• Downsample the majority classes by randomly selecting count of minority class * max_imbalance number of samples from those classes. Reset the index of labels and features if necessary. Throw out the indices that are not in that one
• Reference the lines surrounding np.issubdtype related to bin_count for help determining categorical vs continuous labels

[aiqc]

Separate follow up issue would be using bins to downsample continuous labels