DataTerminology.comNEW
Also known as: MAE
Average magnitude of prediction errors measured as absolute differences from actual values.
DESCRIPTION
Mean Absolute Error (MAE) measures the average magnitude of prediction errors without considering their direction. It uses L1 loss, making it more robust to outliers than mean squared error but less sensitive to large deviations. MAE is expressed in the same units as the target, aiding interpretability. It is convex but not differentiable at zero, which can influence optimization choices. MAE is popular in regression, forecasting, and model monitoring. In robust settings, median predictions minimize MAE. For deployment, MAE complements metrics like MAPE, MSE, and quantile loss to contextualize performance.
ALTERNATIVE DEFINITIONS
scikit-learn
The mean absolute error (MAE) is a risk metric corresponding to the expected value of the absolute error loss.
Connects MAE to risk minimization and L1 loss.
Google Developers (ML Glossary)
Mean absolute error (MAE) is the average of the absolute values of the individual prediction errors.
Emphasizes interpretability as average absolute error magnitude.
RELATED TERMS
Huber Loss L1 Loss Regression Analysis Forecast Accuracy Median Absolute Error Quantile Loss
COMPETING TERMS
KNOWLEDGE AREAS
Mathematics Data Science AI Testing & Validation MLOps
TARGET AUDIENCE
Data Scientist MLOps Engineer AI Engineer Data Analyst
EXTERNAL SOURCES
- Wikipedia - Mean absolute error
- scikit-learn - MAE within regression metrics
- Forecasting: Principles and Practice - Accuracy metrics including MAE
- TensorFlow - Keras MAE metric documentation
TOP QUESTIONS
When should you prefer Mean Absolute Error over Mean Squared Error?
Choose MAE when robustness to outliers is important and interpretability in the original unit is desired. MSE penalizes large errors more heavily and may be preferable for emphasizing outliers.
Category: Alternatives
Why is MAE less sensitive to outliers than MSE?
MAE grows linearly with error magnitude, while MSE grows quadratically. Consequently, large deviations have less disproportionate impact under MAE.
Category: Misunderstanding
What predictor minimizes Mean Absolute Error?
The conditional median minimizes expected MAE. In practice, optimizing MAE tends toward median-like estimates under symmetric loss.
Category: Concept
Is Mean Absolute Error differentiable everywhere?
MAE is not differentiable at zero due to the absolute value. Optimizers handle this using subgradients or smooth approximations like Huber loss.
Category: Requirements
How do you interpret MAE values?
MAE is in the same units as the target and represents average error magnitude. Smaller values indicate better fit, but context and baseline comparisons are essential.
Category: Example
How does MAE relate to L1 loss?
MAE is the empirical mean of L1 losses across predictions. Minimizing MAE corresponds to minimizing L1 risk.
Category: Concept
What are common mistakes when using MAE?
Comparing MAE across different scales without normalization is misleading. Also, relying solely on MAE hides directional bias; complement with signed error metrics.
Category: Mistakes
How is MAE used in model monitoring?
MAE tracks average error drift over time. Establish thresholds and segment by cohorts to detect localized performance degradation.
Category: Management
What is a downside of MAE in optimization?
Non-differentiability at zero can slow convergence for gradient-based methods. Smooth surrogates like Huber loss can help training.
Category: Risks