Why Orange "Predictions" and "Test & Score" produce different results on the same data?

Question

It is not very clear what is the difference between the following two schemes:

From help docs:

• Test & Score widget: tests learning algorithms on data.
• Predictions widget: shows models’ predictions on the data. Which model? Pretrained one? Where is it taken from?

Obviously the results of Confusion Matrix are different. So how should I train and test my model is not clear.

If can easily do this:

What's the point of a separate single Predictions widget connected to some learner? What does it actually test?

Answer 1

Perhaps I can elaborate a bit on this.

Test&Score is used for evaluating a model. You provide T&S with training data and the learner (e.g. Random Forest) and then the widget performs 10-fold cross-validation on training data, leaving out 1/10 of a data for testing. Results of all 10 fold are then combined into a single evaluation result.

Predictions, however, does not perform cross-validation. It doesn't test the data provided. Instead, Random Forest passes a model to Prediction, not a Learner. A model is built on train data. Predictions widget does not consider the real class, it uses the model from RF and applies it to the (new) data.

These are two different procedures, thus you are getting different results.

Answer 2

I guess Test & Score is used to score models on labeled, known data, e.g. via cross-validation, whereas Predictions is used to have trained model make predictions on new, unlabeled data.