weight_mode

Description

weight_mode specifies how to deal with imbalanced datasets.

sample: samples training examples with replacement according to the label weight distribution.
weighted_loss: weighs training examples in the loss function according to the label weight distribution.
mix: samples training examples with replacement according to the label weight distribution, but will re-weigh them in the loss function according to the inverse of the given weight.
The default setting is weight_mode=sample.
Only applicable when majority_sampling_ratio is specified.