First page Back Continue Last page Overview Graphics
Avoid Throwing Away Information
Unlike most Graham-esque filters, CRM114 has no “significance window” of the most extreme N words. Every feature counts, but only a little..
No word or feature can have an overriding impact.
There’s no “ten nonspammy words” that can sneak a spam past the filter.
This totally violates the Bayesian assumption of statistical independence.... but it still works just fine.