Скачать или смотреть Metrics to check classification model | Confusion Matrix | Accuracy | Recall

Accuracy:

·         Definition: The ratio of correctly predicted instances to the total instances.

·         Use Case: Commonly used when classes are balanced.

·         Limitations: Misleading in imbalanced datasets; accuracy may be high even if the model is not performing well on the minority class.

Precision:

·         Definition: The ratio of correctly predicted positive observations to the total predicted positives.

·         Use Case: Useful when the cost of false positives is high.

·         Limitations: Precision alone may not provide a complete picture; should be used in conjunction with recall.

Recall (Sensitivity or True Positive Rate):

·         Definition: The ratio of correctly predicted positive observations to the total actual positives.

·         Use Case: Important when the cost of false negatives is high (e.g., medical diagnoses).

·         Limitations: May not be suitable if false positives are costly.

F1 Score:

·         Definition: The harmonic mean of precision and recall.

·         Use Case: Balances precision and recall; suitable for imbalanced datasets.

·         Limitations: Favors models with similar precision and recall; might not be ideal if you prioritize one over the other.

Specificity (True Negative Rate):

·         Definition: The ratio of correctly predicted negative observations to the total actual negatives.

·         Use Case: Complements recall; useful when the cost of false positives is high.

·         Limitations: Should be used in conjunction with sensitivity; not commonly used in isolation.

Receiver Operating Characteristic (ROC) Curve and Area Under the Curve (AUC):

·         Definition: ROC curve is a graphical representation of the trade-off between true positive rate and false positive rate at various thresholds.

·         AUC: Area under the ROC curve; measures the ability to discriminate between positive and negative instances.

·         Use Case: Useful for binary classification problems.

·         Limitations: Not suitable for imbalanced datasets.