DataTerminology.comNEW
A classification metric defined as the average of recall (sensitivity) and specificity, mitigating bias on imbalanced datasets.
DESCRIPTION
Balanced accuracy evaluates classifier performance by averaging sensitivity and specificity, giving equal weight to each class. It is useful for imbalanced data where raw accuracy overemphasizes the majority class. The metric helps compare models fairly by compensating for class frequency differences. It relates to macro-averaged recall and is implemented in popular ML libraries. Analysts should examine confusion matrices and other metrics alongside balanced accuracy.
ALTERNATIVE DEFINITIONS
scikit-learn
The balanced accuracy computes the average of recall obtained on each class.
Concise formula emphasizing class-wise recall averaging.
imbalanced-learn
Balanced accuracy is the macro-average of recall scores per class and corresponds to (TPR + TNR)/2 for binary classification.
Clarifies relation to TPR/TNR and macro recall.
RELATED TERMS
Recall Specificity Precision ROC AUC Matthews Correlation Coefficient Confusion Matrix
COMPETING TERMS
RELATIONS
- Broader: Accuracy
KNOWLEDGE AREAS
Data Science Testing & Validation AI
TARGET AUDIENCE
Data Scientist Data Analyst MLOps Engineer BI Specialist
EXTERNAL SOURCES
- Wikipedia - Sensitivity and specificity
- scikit-learn - Official metric documentation
- imbalanced-learn - Evaluation metrics for imbalanced data
TOP QUESTIONS
Why use Balanced Accuracy for imbalanced classification?
It weights each class equally by averaging sensitivity and specificity, preventing majority-class dominance in overall accuracy.
Category: Relevancy
How is Balanced Accuracy computed?
For binary tasks, it’s (TPR + TNR)/2. For multiclass, average per-class recall and report the mean.
Category: Requirements
What are common mistakes with Balanced Accuracy?
Treating it as sufficient alone; always inspect confusion matrices and class-wise metrics to understand model behavior fully.
Category: Mistakes
How does Balanced Accuracy compare to F1 score?
Balanced accuracy focuses on per-class recall and specificity, while F1 balances precision and recall. Choose based on error costs.
Category: Alternatives
What is a misunderstanding about Balanced Accuracy?
It is not the same as macro-F1; balanced accuracy averages recall, whereas macro-F1 averages F1 scores across classes.
Category: Misunderstanding
How to improve Balanced Accuracy?
Use class weighting, resampling, threshold tuning, and better features. Evaluate per-class performance to target weaknesses.
Category: Management
Does Balanced Accuracy handle multi-label problems?
It applies per label using recall averages, but specialized multi-label metrics may be more informative.
Category: Opinion
What risks arise when optimizing only Balanced Accuracy?
You may ignore precision, leading to excessive false positives or negatives depending on context. Consider multiple metrics.
Category: Risks
Where to learn Balanced Accuracy usage?
Review scikit-learn and imbalanced-learn docs and study confusion matrices across classes to interpret results properly.
Category: Overview