Confusion Matrix Analyzer

Step-by-Step Guide

1. 1Choose an Input Mode: Select Manual Counts for binary classification with known TP/FP/FN/TN totals, or Actual vs. Predicted Lists for multi-class analysis from raw labels.
2. 2Enter or Import Your Data: Type your confusion matrix counts, paste two label lists (one label per line), or import a CSV file with Actual and Predicted columns.
3. 3Enable Case-Insensitive Matching if Needed: Toggle this on if your labels might differ only in capitalization, such as "Cat" vs. "cat".
4. 4Review the Confusion Matrix Heatmap: See exactly how many predictions fell into each actual-vs-predicted combination, with darker cells indicating higher counts.
5. 5Check the Metrics Dashboard: Review accuracy, precision, recall, F1, and — for binary or two-class data — specificity, MCC, and balanced accuracy, plus a full per-class table for multi-class data.
6. 6Copy or Export: Copy the metrics summary, download the full report as CSV, TXT, or JSON, or print a formatted report.

Key Features

• ✓Binary and multi-class classification support
• ✓Manual TP/FP/FN/TN entry, or raw label list analysis
• ✓CSV drag-and-drop / file import with Actual, Predicted columns
• ✓Automatic unique-class detection for multi-class data
• ✓Interactive confusion matrix heatmap
• ✓Per-class precision, recall, F1, and support table
• ✓Macro and weighted average aggregation
• ✓Binary-only metrics: specificity, NPV, FPR, FNR, MCC, balanced accuracy
• ✓Optional case-insensitive label matching
• ✓Row-count mismatch detection with a clear error message
• ✓Instant calculation with a 150ms debounce as you type
• ✓Export report as CSV, TXT, or JSON, plus a printable layout
• ✓All processing runs locally — no data leaves your browser

Pro Tips

• 💡Use MCC instead of accuracy as your headline metric on imbalanced binary datasets — it accounts for all four confusion matrix cells and is much harder to game.
• 💡Check the per-class table, not just macro/weighted averages, when working with multi-class data — a strong overall F1 can still hide one class performing terribly.
• 💡Use weighted average when your classes are naturally imbalanced and you care about overall performance; use macro average when every class matters equally regardless of size.
• 💡Enable case-insensitive matching if your labels come from different sources that might format the same class differently, like "Positive" vs. "positive".
• 💡Always sanity-check row counts before analyzing — a silent row-count mismatch between Actual and Predicted would otherwise silently misalign every comparison.

Common Mistakes to Avoid

• ✕Don't rely on accuracy alone for imbalanced datasets — a model that always predicts the majority class can score deceptively high accuracy while being practically useless.
• ✕Don't forget that Specificity, NPV, FPR, FNR, and MCC are only computed for exactly two classes — for three or more classes, use the per-class precision/recall/F1 table instead.
• ✕Don't mix up macro and weighted averages when reporting results — they can tell very different stories on imbalanced data, so specify which one you're citing.
• ✕Don't paste mismatched Actual and Predicted lists expecting row-by-row alignment to just work — every row must correspond exactly, in the same order, between both lists.
• ✕Don't ignore the confusion matrix heatmap itself — aggregate metrics can mask specific, systematic confusions between particular class pairs that the raw matrix reveals immediately.