Skip to content

the minimum example per class threshold (2) is very low and at low end cannot be split #75

@realmarcin

Description

@realmarcin

we can increase minimum samples — basically assuming an 80/20 train/test split N>=2 means at the lowest count you get 1.6 in train and 0.4 in test (which is not possible). For robust evaluation we want to maximize the test counts and if you are splitting 80/20 (I think I saw 20%), then if there are 10 examples for the class 8 go to training and 2 to test. But 2 is low … we may want >= 5? which implies min_samples:
0.8/0.2 = X / 5 -> X = 4/0.2 = 20
so if you want at least 5 test examples then the minimum number of total examples for the class should be 25.
this should also improve accuracy since the rare classes are going to be hard to train/predict/evaluate.

@justaddcoffee @turbomam

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions