Two prediction models can have the same average absolute error while making noticeably different kinds of mistakes. One may be consistently a little wrong; another may usually be close but occasionally miss badly. MAE and RMSE help expose that difference, provided you understand what each calculation rewards.
Suppose model A has absolute errors of 2, 2, 2 and 2. Its mean absolute error, or MAE, is 2. Model B has errors of 0, 0, 0 and 8, giving the same MAE: 8 divided by 4.
For root mean squared error, square the errors, average those squares and take the square root. Model A still scores 2. Model B scores 4, because the mean of 0, 0, 0 and 64 is 16, whose square root is 4. The larger miss receives greater influence.
The scikit-learn documentation provides the formal interfaces for mean absolute error and root mean squared error. Both metrics remain in the target variable’s units for this simple single-output case.
Choose the comparison around the job
If an occasional large error is particularly costly, RMSE can help draw attention to it. That does not make RMSE universally better. The application’s consequences may be asymmetric, vary by subgroup or depend on thresholds that neither metric captures alone.
Evaluate models on the same held-out cases and report the target units. Inspect the distribution of residuals and examples of the largest misses. Check whether one model performs poorly for a particular region, product or time period even when its overall number looks acceptable.
Keep the metric calculation separate from the model-training objective. A model can be trained with one loss and evaluated with several measures. The final choice should explain which mistakes matter in the intended use, not merely select whichever metric makes a preferred model look best.