Contact
Confusion Matrix Pro is built and maintained by Isik & Co. Every message reaches a person, not a queue.
How to reach us
Write to misik6@gatech.edu. That address is the single point of contact for everything below, and it is the same one the Send feedback button on the calculator opens. There is no contact form here on purpose: the site has no server and no database, so a form would have nowhere to submit to and nothing to store your message in. A plain email keeps the whole exchange in your own mail client, where you can see exactly what you sent.
We read every note. Expect a reply within a few working days for anything that needs one; a bug report that is already fixed by the time we read it may just get quietly deployed instead.
What to write about
- A wrong number. The highest-priority category. Include the exact TP/TN/FP/FN counts (or the share link, which carries them), the metric, and what you expected. Every binary metric is covered by a golden-value test suite, so a discrepancy is either a real bug or a definitional disagreement worth having.
- A bug. Broken export, a control that does nothing, a layout that collapses. Your browser and operating system help.
- A feature request. Especially a metric or confidence-interval method you need and cannot find.
- Citation and licensing. The source is MIT-licensed and the tool is free to use, cite, and embed. See the Cite tab on the calculator for ready-made APA and BibTeX entries.
- Data and privacy questions. Answered in full under About & disclaimer, but ask if anything is unclear.
What we can't help with
We can't interpret your specific model results, advise on clinical or financial decisions, or debug your training pipeline. The tool computes metrics and shows their formulas and confidence intervals; what those numbers mean for your problem is yours to judge. Results are provided for informational and educational purposes only, with no warranty as to the accuracy of any calculation and no liability for decisions arising from their use.
For a question about what a metric means, the metric glossary and the formula reference are usually faster than an email.
Before you write
Will my data be seen if I send a share link?
A share link encodes the counts or predictions you entered into the URL itself. If you paste one into an email, whoever reads that email can see those values. Nothing about your session is transmitted to us otherwise: every calculation, including CSV uploads, runs locally in your browser and is discarded when you close the tab. If your data is sensitive, reproduce the bug with substitute numbers and send those instead.
My numbers disagree with scikit-learn. Is that a bug?
Check the matrix ordering first. This tool's binary matrix is positive-first, with TP in the top-left. sklearn.metrics.confusion_matrix returns [[TN, FP], [FN, TP]] instead. Reading one as the other swaps sensitivity and specificity and changes precision and NPV. The import feature assumes sklearn's order and says so in the interface. If the ordering matches and the numbers still disagree, that is worth an email.
Why is my confidence interval different from the one I calculated?
Because there is more than one correct method. Proportions here use the Wilson score interval rather than the normal (Wald) approximation, which is unreliable at small counts and near 0 or 1. Ratio metrics (LR+, LR−, and the diagnostic odds ratio) use the log-normal method of Simel and Altman, since those quantities are unbounded above. Linear combinations of two proportions, such as Youden's J and markedness, use a normal approximation. Each metric card names its own method.