Recall

Recall is the ratio of captured true positives.

Recall

Recall measures how many of the true positives were caught: TP / (TP + FN). It matters where missing is costly: disease screening, fraud detection.

from sklearn.metrics import recall_score
y_true = [1, 1, 1, 0, 0]
y_pred = [1, 0, 0, 0, 0]
print(recall_score(y_true, y_pred))   # 1 / 3 = 0.333

A model that calls everything positive gets a recall of 1.0, but its precision drops a lot.

The precision-recall trade-off

When the threshold drops the model says positive more often: recall rises, precision falls. When it rises, the opposite happens:

p = m.predict_proba(X_test)[:, 1]
for esik in [0.3, 0.5, 0.7]:
    t = (p >= esik).astype(int)
    print(esik, precision_score(y_test, t), recall_score(y_test, t))

Which point to choose depends on the business problem; precision_recall_curve gives all thresholds at once.

Exercise Threshold effect

For the probability array p, produce decisions at the 0.3 and 0.7 thresholds. For each threshold print the threshold, the precision and the recall with round(..., 2) on the same line.

Show example solution

import numpy as np
from sklearn.metrics import precision_score, recall_score

y_true = np.array([1, 1, 1, 1, 0, 0, 0, 0])
p = np.array([0.9, 0.8, 0.45, 0.35, 0.6, 0.4, 0.2, 0.1])

for esik in [0.3, 0.7]:
    t = (p >= esik).astype(int)
    print(esik, round(precision_score(y_true, t), 2), round(recall_score(y_true, t), 2))

Test yourself

In cancer screening, which metric should be kept high first?
- Precision
- Accuracy
- None of them
- Recall
Show answer

Recall — Missing a patient (FN) is much costlier than sending a healthy person for an extra test.

Compute recall.

r = ____(y_true, y_pred)

Show answer

r = recall_score(y_true, y_pred)

Match the threshold change with the results.

Threshold drops

Recall rises
Threshold rises

Precision rises
Positive for everything

Recall 1.0
precision_recall_curve

All thresholds

Next lesson: F1 score

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies