ITPEC IP April 2025 Question 3
The automation of the inspections performed by a quality controller is considered. By providing as training data 10,000 product images and the results of judgments by the quality controller about whether each product is defective or not, a machine learning model is created to judge whether a product is defective or not. The results of a test judgement by a machine learning model on 100 product images are as shown in the table. When the ratio of the number of images that product is judged to be defective by the machine learning model out of the images that product is judged to be defective by the quality controller is deemed to be the recall rate, what is the recall rate in this test judgement?
Explanation
0.50 — recall takes the images the quality controller judged defective as its base and asks what share of them the model also caught.
Read the table by rows, since the base is the quality controller's judgment:
- •Defective by the quality controller: 5 caught by the model + 5 missed = 10 images
- •Of those, judged defective by the model: 5
- •Recall: 5 ÷ 10 = 0.50
The 15 images in the lower row are false alarms: the model called them defective, but the quality controller did not, so they do not enter recall at all.
Answer (c)
Why not others:
- •(a) 0.05 — 5 ÷ 100: divides the caught defectives by all 100 test images instead of by the 10 truly defective ones
- •(b) 0.25 — 5 ÷ (5 + 15): takes the model's Defective column as the base; that is precision, the share of the model's defective verdicts that are right
- •(d) 0.80 — (5 + 75) ÷ 100: the accuracy, the share of all images on which the model and the quality controller agree
Key rule: Recall = true positives ÷ all actual positives (how many defects are found); precision = true positives ÷ all predicted positives (how many alarms are real).
Basic Theory · AI-written, may contain errors
This question comes from an official ITPEC past paper. ITPEC Practice is an independent study tool and is not affiliated with ITPEC. See the official IP past-paper collection or Report an issue.