The entropy gain of the class distribution by the model over the prior distribution (see class_complexity_gain), divided by the number of instances in the input data.