Skip to content

PHIlter Corrected Precision

beaunorgeot edited this page Sep 18, 2018 · 4 revisions

This page tracks manually updated precision values during Philter development, by dataset. Updated precision values are based on the assumption that certain "error" false positives are the result of expected Philter behavior.

UCSF Batch 5

Date: 09/17/18, Test: Whitelist-only

Initial Whitelist-Only Performance

TPs: 1791
FPs: 1137
Precision: 61.17%

Error FP Counts By Category:

Words in non-English language: 316
Word concatenations: 146
Misspellings: 45
Names: 66
Locations: 58

Total: 631
Proportion of FPs: 55.50%

Uncorrected Whitelist only Performance

Recall: 26.56%
Precision: 70.49% (up from 61.17%)
Retention: 99.09%

Manually Corrected Whitelist only Performance

TPs: 2422
FPs: 506
Precision: 82.72%