Explainable AI, online
Ask any model why: which inputs pushed its answer up or down, by how much, and what would have to change.
For one prediction
SHAP values on your original columns - categories included - that add up exactly to the prediction, with the reason in plain words.
A counterfactual: the smallest change that flips the answer, and the most similar real example that got the other answer.
For the whole model
Permutation importance, partial dependence (how the prediction moves as one input changes), SHAP summaries over many rows - and for text and images, the words or pixels that drive each class.
Try it on mushrooms: the model learns from 22 descriptions, and the explanation shows it relies mostly on the smell - just as mushroom guides do.
Try it in a minute
- Press Try it now - mushrooms, preprocessing and a trained network are set up.
- Read why it called a mushroom poisonous.
- Open The whole model to see what it relies on overall.
Questions
Which models can be explained?
All of them - neural networks and classic ML (trees, forests, boosting, SVMs...).
Are the SHAP values exact?
For small inputs they are computed exactly; otherwise they are sampled and still add up exactly to the prediction.
Does it work for text and images?
Yes - words that push each class, and image patches or integrated gradients for pictures.