Metric evaluation error: y_true and y_pred contain different number of classes

[nicl-nno]

The problem "Metric evaluation error: y_true and y_pred contain different number of classes 45, 85" raises for click_prediction_small dataset

Looks like class stratification is failing.

[curiousHG]

Can you expand on the issue? Or maybe narrow down to which file or code the problem is?

[nicl-nno]

@curiousHG

This problem occurs for classification tasks when the number of classes is very large and cross-validation fails to preserve the examples for all classes.

The connected problems are for <2 classes in sample and new classes in predict sample.

[gkirgizov]

Duplicate of #954