Statistical Content Filtering
Statistical content filtering, often known as "bayesian filtering" was first proposed by the tech visionary Paul Graham in a piece entitled, "A Plan For Spam". By doing a statistical analysis of the text contents of a spam message it's possible to determine the likelihood of it being spam. In order for this to work, the filter must be trained, being fed samples of good mail (sometimes known as ham) and spam mail. Once it has been fed a sufficient amount of sample mail (usually at least 500 messages of each, but it depends on the individual filter), it can begin classification with an exceptionally high degree of accuracy. Because one user's spam may be another user's desired email, statistical filters are hard to deploy on a site or enterprise wide basis.
Currently, some email clients, notably Mozilla, its derivative Thunderbird, and Apple's Mail.App have been shipping with statistical filtering built into their mail clients. It appears that Outlook 2003 also contains a statistical filtering as part of its "Junk Mail Filters" feature. For those not using the newest version of Outlook, there is an Open Source plug-in called SpamBayes for most versions of Outlook, adding statistical filter capabilities to this very ubiquitous email program.
Collaborative Filtering
Collaborative filtering leverages the power of the human brain to classify spam and the power of the Internet to share those classifications. In general, these require the use a subscription service. When a user, call her Alice, receives a spam email, she classifies it as spam. The collaborative filtering software then generates what is known as a checksum, an unique ID computed from the content of the message. This checksum and the spam classification is then transmitted over the Internet to a "checksum clearinghouse", the heart of the service. Later, when another user, call him Bob, receives the email, his version of the software looks up Alice's classification using the checksum. Bob's email program can then safely classify the email as spam without his intervention. In order to make a collaborative system more robust, different trust metrics and ratings are overlaid on submissions to ensure better results.