Nadir,

The following is an excerpt from the readem.html file in the SpamFilter directory, which tells how the Bayesian filtering will train itself. If the corpus floder is empty, please check that the box labeled "learn new incoming emails" under the Settings - bayesian Filter tab is checked.

Roberto F.
LogSat Software

Bayesian Statistical Filtering
The new v2 release of SpamFilter ISP features statistical DNA fingerprinting of incoming emails. The statistical analysis is performed using Bayesian rules. Tokens within incoming emails are scanned and categorized in a corpus file. The content of all new incoming email is fingerprinted and checked against the historical data. If there is a high statistical probability that the email is spam, it is rejected.

The statistical engine kicks in after 5,000 non-spam and 5,000 spam emails have been received (values customizable by editing the SpamFilter.ini file). This is done to build a valid statistical base to use before emails are rejected. During this period of time, it is critical to avoid false positives. If a good email is quarantined, forcing it's redelivery either thru the web interface or the SpamFilter GUI will "teach" SpamFilter that the fingerprint in that email is a "good" one, and the statistical DNA database will adapt itself to it. It is very important initially to check the quarantine often to force delivery of legitimate email that has been blocked by the "regular" filtering rules.

A slider is used to control the accuracy of the statistical filter. Incoming emails are assigned a probability of being Spam, ranging from 0% (most likely a valid email) to 100% (most likely Spam). Any emails that have a probability of being spam above the value you set will be rejected. Typical threshold values are in the 99.9% range.

Edited by LogSat - 24 January 2009 at 10:12am