Bill Yerazunis: Better than Human

(Bill Yerazunis' CRM114 has the best spam filtering performance I've heard of to date and deserves to be better known. According to this email, which Bill has allowed me to reprint, CRM114 now achieves 99.87% accuracy. This level of performance is more evidence that probabilistic, content-based filters are the answer to spam.)

Date: Wed, 16 Oct 2002 10:07:10 -0400
From: Bill Yerazunis

The current statistics for CRM114's sparse binary polynomial hashes are in. Over the last two weeks it's managed a reasonably enviable 99.87% accuracy.

Details: 2374 fresh-from-the-wild incoming messages, 1518 spam and 856 nonspam yielded only three errors. All classification was done via the sparse binary polynomial hash classifier system, no blacklists or whitelists used, training set approximately 250Kbytes. If I'd turned on the whitelists and blacklists, it would have been one error instead of three, for a 99.96% accuracy, but as we consider blacklists (even of known spamhausen) to be "cheating", I'm running without them.

The worst-case margin of error on that result is +-0.16%, as that's the measured error rate of _me_ distinguishing spam from nonspam. (I manually classified the same set of 1900 messages twice, and found three errors in my own classifications, hence I have a 99.84% success rate.)

With an accuracy of 99.87% for CRM114 versus my own rate of 99.84%, I conclude that CRM114 is "better than human", or at least better than its creator when confronted with 3800 decisions to make.

Date: Wed, 16 Oct 2002 14:26:46 -0400
From: Bill Yerazunis

[The three errors were] one false reject, two false accepts, if I recall correctly.

But only one of the false accepts was from an unknown source, and the false reject was from a known-good source, so the black/whitelist would have caught it, and it would still be a 99.96% accurate system.

More Info:

Bill Yerazunis
Plan for Spam