Scores, Probabilities, Labels, and First Thresholds

Unit ID: ML-M03-U05 Estimated active time: 25-35 minutes

Why this matters

A classifier often gives more than a final label.

It can also give a score or predicted probability.

The label is the decision the model reports. The probability is a measure the model uses before that label is chosen.

The idea

For a binary classifier, the model may estimate:

probability of class 1 = 0.73

If the default threshold is 0.50, then 0.73 becomes label 1.

If the probability is 0.31, it becomes label 0.

This does not mean the model knows the truth. It means the model estimated a score from the learned pattern.

Predict

With a default threshold of 0.50, what label comes from each probability?

Answers: 1, 0, 1.

Run or inspect

In scikit-learn, many classifiers support:

predicted_probabilities = model.predict_proba(X_test)

For binary classification, this returns probabilities for class 0 and class 1.

In the first notebook, we inspect a few class-1 probabilities so learners can see how labels are created.

Change one thing

If the cost of missing a struggling learner were much higher than sending one extra support message, we might choose a lower threshold later.

But threshold tuning is not a Module 3 topic. It gets proper time in Module 6.

Practice

In the worked notebook:

1. Inspect a few predicted probabilities.
2. Convert them to labels with the default threshold.
3. Compare the labels with the true target.
4. Write one sentence about why probability is not certainty.

Check and explain

Complete:

A predicted probability is useful, but it is not ______.

Takeaway

Classifiers can produce probabilities and labels. Thresholds turn scores into labels, but threshold decisions need careful context.