Analysis Results
Interactive Confusion Matrix
Click a heatmap cell to inspect matching records.
Per-Class Performance
Actual and Predicted Distribution
Per-Class Metrics
Most Confused Class Pairs
Confidence Summary
Misclassified Images
Model Comparison
Interpretation
Formula Used
| Accuracy | (TP + TN) / (TP + TN + FP + FN) |
|---|---|
| Precision | TP / (TP + FP) |
| Recall | TP / (TP + FN) |
| Specificity | TN / (TN + FP) |
| F1-score | 2 × Precision × Recall / (Precision + Recall) |
| Balanced accuracy | Average recall across classes |
| Cohen’s kappa | (Observed accuracy − Expected accuracy) / (1 − Expected accuracy) |
| Multiclass MCC | Generalized correlation between actual and predicted labels |
How to Use
- Paste records or upload a CSV dataset.
- Choose filters, confidence rules, and class settings.
- Select normalization and heatmap display preferences.
- Press Calculate and Plot to generate results.
- Inspect matrix cells, errors, metrics, and confidence charts.
- Export tables, graphics, settings, or a PDF report.
Example Data
| Image | Actual Class | Predicted Class | Confidence |
|---|---|---|---|
| image_001.jpg | Cat | Cat | 0.96 |
| image_002.jpg | Dog | Cat | 0.71 |
| image_003.jpg | Bird | Bird | 0.91 |
| image_004.jpg | Cat | Dog | 0.64 |
| image_005.jpg | Dog | Dog | 0.88 |
Frequently Asked Questions
What does the diagonal show?
Diagonal cells contain correct classifications for each image class.
What do off-diagonal cells show?
They show which actual classes were confused with other classes.
Which normalization should I use?
Normalize by actual class when comparing recall across unequal classes.
Why can accuracy be misleading?
Large classes can dominate accuracy while minority classes perform poorly.
What is macro F1?
It averages class F1-scores without weighting by class support.
What is weighted F1?
It weights each class score using its number of samples.
How are missing confidence values handled?
They remain available unless the keep-missing option is disabled.
Can I compare multiple models?
Yes. Include a model column and keep the model filter empty.
How are rejected predictions treated?
Predictions below the threshold become the configured uncertain class.