I have long relied on CRM114 for spam filtering on my mailserver, and have generally been fairly happy with it. But it never hurts to try some alternatives, so I tried some lighter-weight programs: namely QSF and BMF. This was not a full-fledged test, mind you, just a report of my personal results.
The general setup is as follows:
QSF, BMF, and CRM114 called in serial from procmail. CRM114 was last in the list to catch anything the first two missed, since it was the most “experienced”. Emails to all three programs were piped through normalizemime, which deobfuscates emails and also removes headers that could skew the result. Finally, after feeding a single ham & spam email to the programs, all training was done “on error” (TOE). Maybe this is not optimal for these programs?
Results
Although both QSF and BMF are fast and lightweight, neither had the accuracy of CRM114. Moreover, the errors they made seemed to be “sticky”, in that retraining didn't correct the issue even when receiving multiple similar emails from the same source. After several weeks, both were still missing the same types of spam and still mis-tagging the same non-spam messages.
I'll be sticking with CRM114. Besides its accuracy, the “reaver_cache” also stores the emails making retraining much easier and less error-prone. Your mileage will vary, but I'm again convinced it is one of the best filters out there.
