Differences between revisions 3 and 4
Revision 3 as of 2007-02-17 03:43:34
Size: 1643
Comment:
Revision 4 as of 2009-09-20 23:17:10
Size: 1643
Editor: localhost
Comment: converted to 1.6 markup
No differences found!

FP%/FN% Percentages

The main system used to measure spam-filtering accuracy in SpamAssassin is the "FP%/FN% percentages" system.

It's quite simple. First, you scan a corpus of hand-classified mail (see HandClassifiedCorpora) to get 4 figures:

  nspam   = number of known-to-be-spam messages in the corpus
  nham    = number of known-to-be-ham (nonspam) messages in the corpus
  fp      = number of ham messages incorrectly marked as spam
  fn      = number of spam messages incorrectly marked as ham

fp is so named because it's more commonly and concisely called a FalsePositive, and fn a FalseNegative.

Next, perform this calculation:

  FP% = (fp / nham) * 100
  FN% = (fn / nspam) * 100

and you have two numbers that simply, concisely, and comprehensibly describe the accuracy and performance of the filter.

For example, let's say we do a test as follows:

  nspam   = 1000
  nham    = 1500
  fp      = 2
  fn      = 30

the FP% and FN% work out as (2 / 1500) * 100 = 0.1333% and (30 / 1000) * 100 = 3.0% respectively.

The 'STATISTICS.txt' files distributed with SpamAssassin versions since about 2.30 include this data, measuring the ruleset's accuracy against a validation ruleset:

# SUMMARY for threshold 5.0:
# Correctly non-spam:  29443  99.97%
# Correctly spam:      27220  97.53%
# False positives:         9  0.03%
# False negatives:       688  2.47%
# TCR(l=50): 24.523726  SpamRecall: 97.535%  SpamPrec: 99.967%

As you can see, FP% and FN% get pride of place in the measurement scheme.

See also MeasuringAccuracy for other methods.

FpFnPercentages (last edited 2009-09-20 23:17:10 by localhost)