Owners: Sam, Unihedron, Patrick Hofman, Jan Dvorak & ProgramFOX. Total terms: 939. Accuracy threshold: 17.5%. Full scan enabled: True. Posts caught over last 7 days: 546. Uptime: 00:10:08.3857976.
Owners: Sam, Unihedron, Patrick Hofman, Jan Dvorak & ProgramFOX. Total terms: 939. Accuracy threshold: 17.5%. Full scan enabled: True. Posts caught over last 7 days: 553. Uptime: 00:11:59.1971357.
@Jan Regarding #28. I've just taken a second look and realised what you actually said, "sensitivity (true positive / all bad) and specificity (true negative / all good)" (and not, "total tp/total fp" and vice versa). I'm not quite sure what you mean by "true positive" and "true negative" (as in, I'm not exactly sure which metrics you're referring to).
Owners: Sam, Unihedron, Patrick Hofman, Jan Dvorak & ProgramFOX. Total terms: 943. Accuracy threshold: 17.5%. Full scan enabled: True. Posts caught over last 7 days: 731. Uptime: 01:33:40.6566456.
@Sam tp = true positive = bad posts detected as bad posts. fp = false positive = good posts detected as bad posts. tn = true negative = good posts detected as good posts. fn = false negative = bad posts detected as good posts
there would be a separate sensivity metric for spam and for LQ
the number of good and bad posts is the same for all keywords (well... not quite; you start measuring at different times), but different keywords have a different amount of positives
it does make sense to reset all stats after each keyword manipulation
Right, so when someone TP's a report, I'll add the black terms to a "TP" list (or increment some val indicating how many times the term has been TP'd). And the same FP'ing (obviously an "FP" list will be used instead).
Which I can then use in the term-specific stats calculations.
or perhaps distribute unhandled reports between TP and FP proportionally instead? (same results, but the math might be simpler - if 50% handled reports are true, assume 50% unhandled reports are true as well)
