>Has anyone seen anything about this .. Any good or bad ? .. Bayesian >filtering .. Check this link http://www.paulgraham.com/filters.html > >Scott are you thinking along any of these line for the future of >Declude?
Here's part of an E-mail that I posted previously: --- The concept is flawed, but sounds like it could work well if implemented properly. It's essentially the same thing that the heuristics test (which never made it out of beta) does, except that the heuristics test only looks at about 100 different aspects of the E-mail (not necessarily individual words). The flaw is that it is based on the "Bayes Rule", but doesn't meet its conditions (the individual probabilities used must not be related to each other, but they are in this situation). For example, if you have 10 spams that each have the term "This is a spam" in them, and it's determined that each word has a 99% chance of indicating spam (IE if the word "This" appears in an E-mail, there's a 99% chance it is spam), the Bayes Rule would say that an E-mail with "This is a spam" would have about a 99.99999% chance of being spam (when in reality, I believe it would only have a 99% chance of being spam). This probability stuff gets pretty complex unless you've taken a probability and statistics course recently (and did well in it). Another problem is that you need a good database of legitimate E-mail to work with, which is harder to obtain than a good database of spam (which isn't always easy to get). If you just use E-mail to a specific person, it will skew the results. We are doing some testing here to see if the concept could work well (as it seems to for Paul Graham). -Scott --- [This E-mail was scanned for viruses by Declude Virus (http://www.declude.com)] --- This E-mail came from the Declude.JunkMail mailing list. To unsubscribe, just send an E-mail to [EMAIL PROTECTED], and type "unsubscribe Declude.JunkMail". The archives can be found at http://www.mail-archive.com.