Trainable vs. Blacklist-Based

AGAVA SpamProtexx is trainable. Some users run away from trainable spam filters, because they believe that training is a long and tiresome process, which leads to filtering based on obscure and fuzzy criteria. These users prefer filters based on blacklists such as SpamCop, ORDB, DSBL, SPEWS, etc.

Myth One. A blacklist is an objective criteria. Let us describe to you in a few words how a typical blacklist operates. Say, there’s a web hosting company that sells email services. And say, there’s this client, who decides one day to send out a bunch of spams to a lot of addresses. Some of these might get submitted to a blacklist as spam. What happens is the web host’s IP address (even a group of IP addresses usually) gets blacklisted. As a result, you stop receiving legitimate email from the network neighborhood.

You might say: bad web hosting company, bad clients. Well, that might be true, but its not the point. While the bad software company investigates its problem with one bad client, a whole bunch of good and innocent clients would suffer. Is that objective enough? We don’t think so.

Myth Two. Statistics-based criteria is not objective. Roughly, a statistics-based filter analyses each word that it finds in an email and compares it to the entry of the same word in its database, where there’s a spam coefficient for it. The word-based approach has had a number of poor implementations, which might have spoiled the perception of this method. These filters came with a number of pre-defined words/coefficients and raised the spam probability of a message by simply incrementing a spam score of the message. The items that increased the spam score could be removal from list instructions, non-white HTML backgrounds or such words as “free” or “spam”. These filters were known as “newsletter killers”, because if there are two things that are present in any legitimate newsletter, they would be removal instructions and ads for free giveaways. Newsletter publishers went mad and started using tricks to fool these filters by typing “sp*m” or “fre*e”. Well, that is subjective, but there’s a great difference in how SpamProtexx works.

First, the formula that calculates the total spam score is not that simple (not linear). It does not work by simply adding up the coefficients or increasing a spam score, but deploys other criteria, such as the overall message size — if there’s a dozen spam words in a long email about cooking, SpamProtexx will not mark it as spam.

Second, SpamProtexx is trainable. Not configurable, but trainable. That is you cannot tune the coefficients manually, but you can only submit a full message for training. This ensures objectivity by removing human error and also means that your statistical coefficients are based on your personal email.

Both of the above ensure that SpamProtexx’s approach is objective and is capable of providing good results.