Unattended spam filtering using machine learning: implementation, deployment and lessons
John Graham-Cumming
polymail is a commercial anti-spam library in wide use worldwide. The library supports both
attended (i.e. user gives feedback on spam filtering mistakes - sometimes known as 'train on
error') and unattended (i.e. blackbox with no user interaction necessary) filtering.
This talk details the implementation of 'train on everything' spam filtering with SURBL integration
and automatic whitelisting, the deployment of an unattended spam filter and lessons learned
in over a year of deployment supporting 10,000s of clients.
Although polymail is a commercial product, this is a technical talk. Come prepared for
some mathematics, lots of code, plenty of real-world data and no wild marketing claims
of 99.999% accuracy!
