Classification Model Evaluator
Evaluate your binary classifier using real predictions, explore decision thresholds, and quantify the cost of errors. Your data stays in your browser.
Performance overview
Confusion matrix
Rows are actual classes; columns are predicted classes.
Discrimination curves
ROC curve
Precision–Recall curve
Threshold comparison
Illustrative thresholds plus the current selection. Optimize on validation data, not your final test set.
| Threshold | Precision | Recall | F1 | FP | FN | Total cost |
|---|
Interpretation guide
Precision vs. recall
Precision measures the fraction of predicted positives that are correct. Recall measures the fraction of actual positives found. Lowering the threshold often improves recall while increasing false positives.
Accuracy and class imbalance
A classifier can achieve high accuracy by predicting the majority class. Compare precision, recall, specificity, balanced accuracy, and prevalence.
ROC-AUC vs. Average Precision
ROC-AUC measures ranking performance across false-positive rates. Average Precision summarizes precision across recall increments and is often more informative for rare positive classes. AP depends on class prevalence.
Threshold optimization and data leakage
Choose a threshold using training/validation predictions and a predefined objective. Reusing the test set for threshold selection biases reported results. Cross-validation or a dedicated validation set helps avoid leakage.