Bayesian antispam filtering

Bayesian (Statistical) filtering is the way of filtering out the e-mail letters containing spam letters, based on identifying and sorting them with certain indications. The filter examines the letter's content (the headers, single words and word phrases, HTML code) to define spam and compares the e-mail by the known tokens to identify and to sort incoming letters. Then, it calculates the percentage of spam and legitimate words from the original data base it has and defines if the letter is spam or not.

Except the abilities of above-mentioned data, the Bayesian filter can also be trained to define new tokens and accept new user's parameters of identifying spam. And another feature: instead of using only two categories - spam and allowed letters, the filter has the third one - «unsure», which has to be checked by the owner of the e-mail using Bayesian filter. It seems to be an extra work for the user, but it safes from marking useful letters as spam.

But the filter is not just can be taught – is has to be, because it's original data base is sufficiently small and at first, the user has to mark letters as spam manually. From the one side, the trainee of filter is rather handy, but from the other side, a big amount of users would like to get no spam e-mails at all, from the very beginning. Besides, the filter can miss long spam letters, containing a lot of neutral words – it can be, for example someone's long life story with following request to send somewhere a certain amount of money or some proposition like buying an offered commodity.

The Bayesian filtering is rather young and has certain disadvantages. But, comparing with other filters, it has a reliable system of protection against false positive and false negatives and, accordingly, it has lower indexes by these parameters. The filter doesn't let the users pass useful e-mails and, after a certain time of teaching the filter, it blocks almost all the incoming spam.

Despamious uses Bayesian filtering as a part of its antispam protection scheme.