Teaching machine

Synthetic Entity Resolution, Live

A good matcher is not enough. Point one at a population where real matches are rare, and most of what it flags is innocent. Tune the matcher and the base rate, and watch precision, the false-positive burden, and the cost per true match move against each other.

Match threshold 0.55
← loose, catch everythingtight, miss links →
How rare a real match is 1 in 1,000
← rarermore common →
Economics inputs
Precision 2% of flags that are real
Recall 90% of real links caught
False positives 50,000 innocent people flagged
Cost / true match $920 the number that bites

Of every 1,000 alerts

Who is actually in the pile your analysts have to work.

19 real 981 innocent

Of every 1,000 alerts, 19 are real. 981 are ordinary people.

What this shows. This is base-rate neglect, the mistake Tversky and Kahneman named in 1974. Accuracy in the brochure does not survive contact with a low base rate. Drag the base rate rarer and watch the pile fill with the innocent.

The matcher's trade-off

Loosen to catch more, and precision collapses. Tighten to stay clean, and you miss real links.

Precision Recall False positives (shape)

What this shows. Better resolution is not a nicety. It is what makes monitoring both affordable and fair, because it moves this whole curve, not just your place on it.

All data is synthetic and the model is illustrative. A teaching model of a public, decades-old technique, not a description of how any live system works.

Paired with the article The 80,000-Entity Wave.