Several metrics in scikit-learn are using np.average(..., weight=sample_weight) under the hood:

• mean_absolute_error
• mean_squared_error
• explained_variance_error
• r2_score
• mean_tweedie_deviance

When computing the average, numpy will divide by the sum of the weights. Is it the intended behaviour. For instance:

sample_weight = [1, 2, 3, 4]

and

sample_weight = [1/10, 2/10, 3/10, 4/10]

will lead to the same error/score. Dividing by the sum of the weight will also remove any meaning about the use of a unit (if sample_weight is related to a business unit for instance).

So I was wondering if we should multiply the average by the sum of the weight or not.