Chapter 2 · Machine learning basics
Accuracy, precision and recall
Accuracy can mislead; precision asks 'how many flagged were right', recall asks 'how many right ones were found'.
Chapter 2 · Machine learning basics
Accuracy can mislead; precision asks 'how many flagged were right', recall asks 'how many right ones were found'.
The net catches every item to its right and calls it a fish. Watch the four counts and the three scores change as it slides left.
For each item it gives a score for “fish”. Above the threshold it says fish; below, boot. The threshold is the net's edge.
Drag the threshold. Green rings are fish caught, amber rings are boots caught by mistake, dashed magenta rings are fish missed.
A middle setting: a fair balance between boots caught by mistake and fish missed.
A simplified teaching version: a fixed, made-up pond and a pretend classifier score for each item.
Only 1 in 100 emails is spam. A filter that calls everything "not spam" has 99% accuracy. What does it have for recall of spam?
You lower the threshold so the net catches more. Recall usually…
Which metric answers "of the items I flagged, how many were truly positive?"