F1-Score

The F1-score (also known as the F-measure) is a single-number performance metric that provides a balanced summary of a Classification model's success by combining Precision and recall.

📊 See also: Accuracy, Confusion Matrix, Area Under the ROC Curve (AUC), Cohen's Kappa statistic It is especially valuable for evaluating models on imbalanced datasets where Accuracy can be highly misleading.

1. Calculation and Formula

The F1-score is defined as the harmonic mean of Precision and recall.

• Standard Formula: F1=2×Precision+RecallPrecision×Recall​.
• Confusion Matrix Formula: Using counts of True Positives (TP), False Positives (FP), and False Negatives (FN), it is calculated as:

2. Logic of the Harmonic Mean

Unlike a standard arithmetic average, the harmonic mean gives much more weight to low values.

• Requirement for Excellence: A classifier will only achieve a high F1-score if both Precision and recall are high.
• Sensitivity to Extremes: If one metric is perfect (1.0) but the other is very low (e.g., 0.0001), the arithmetic mean would be roughly 0.5, while the F1-score would be near zero. This prevents "gaming" the metric by predicting only the most frequent class or identifying only one positive instance perfectly while missing all others.

3. Why the F1-Score is Important

• Single-Number Evaluation: During model development, teams often try many architectures and parameters. Having a single-number metric like F1 allows them to sort models and quickly decide which ideas are working best without struggling to compare two separate values (Precision and recall).
• Addressing Class Imbalance: In a scenario where the positive class is rare (e.g., 1 in 1,000,000), a "dumb" model that always predicts "negative" will have near-perfect Accuracy but zero recall and an undefined or zero F1-score, correctly identifying it as useless.
• Trust and Reliability: It captures a model's "trustworthiness" (Precision) and "inclusivity" (recall) simultaneously, making it a better proxy for human intuition of a "good" model than Accuracy.

4. Interpretability and Limitations

• Difficulty of Interpretation: A primary disadvantage is that F1-scores are harder to explain and interpret to stakeholders than simple Accuracy.
• Asymmetry: The F1-score is an asymmetric metric; its value changes depending on which class is declared the positive class. For example, if a model for cancer detection has an F1 of 0.17 for the "cancer" class, it might have an F1 of 0.95 for the "normal" class.
• Single Operating Point: The F1-score only captures performance at one specific decision threshold (usually the default 0.5). It does not provide the same detailed insight as a full Precision-Recall Curve, which shows tradeoffs across all possible thresholds.

5. Multiclass Variations

To evaluate multiclass problems, binary F1-scores are computed for each class and then averaged using different strategies:

• Macro Averaging: Computes the unweighted average of per-class F1-scores, treating all classes as equally important regardless of their size.
• Weighted Averaging: Computes the average of per-class F1-scores weighted by support (the number of true instances in each class), making it more representative of the overall dataset distribution.
• Micro Averaging: Computes total TP, FP, and FN across all classes first, and then calculates a single F1-score. This is recommended if you care about each individual sample equally.

6. The Generalized F-Measure (Fβ​)

In many contexts, Precision and recall are not equally important. The generalized Fβ​ score allows practitioners to weight them differently.

• Formula: Fβ​=(1+β2)×(β2×Precision)+RecallPrecision×Recall​.
• F2​ Score (β=2): Weighs recall twice as high as Precision, useful when missing a case is very costly (e.g., shoplifter detection or medical screening).
• F0.5​ Score (β=0.5): Weighs Precision twice as high as recall, useful when false alarms are dangerous (e.g., kid-safe video filters).