Meaning of negative/positive instances for scale_pos_weight

[kawaho]

Hi,
I want to understand what negative/positive instances in the documentation mean for the use of scale_pos_weight. Some resources online said “negative instances” mean the class with smaller number of instance while others said it simply means the class with label = 0. Which one is correct?
Also, for the case where samples are weighted, I assume sum(negative instances) / sum(positive instances) would mean “weighted instances”?
Thanks,
Ka Wa

[hcho3]

Here, the negative instances are the ones with label = 0.

@kawaho

sum(negative instances) / sum(positive instances) would mean “weighted instances”?

Yes, this is a typical value for scale_pos_weight. From https://xgboost.readthedocs.io/en/latest/parameter.html:

A typical value to consider: sum(negative instances) / sum(positive instances)