Not All Bayesian Filters Are Created Equally
I have been using Bayesian filters for my email for a while now. I have tried several different ones, and I have had varied experiences with them. Due to a change in client usage, I had to abandon some in favor of others, but in the end I have benefitted from more knowledge and, ultimately, better solutions.
I used to use The Bat! for my email here in the office. It does a wonderful job of sorting mail and even has a Bayesian filter plugin included, called BayesIT! that supposedly does a good job of chopping down spam. That was not my experience at all. In fact, BayesIT! never once successfully filtered anything, even after training. I switched to Winkler's Bayes Filter Plugin and got significantly better results, with a success ratio around 95%. That's not so bad, but it still means about 50 spams in the inbox per thousand messages.
I have since been converted over to Outlook here in the office (required, due to the calendar and coordination tools) and again I found myself looking for a good filter. I have SpamEaterPro which is a heuristic filter, but it's accuracy has a lot to do with the diligence of maintaining the whitelists and blacklists and it has gotten squirrely in its stability. Outlook's built-in spam filter is ok, but it still is not as effective as what I had before. So, I turned it off and looked for a better way. After trying several filters, I found a freebie on SourceForge called SpamBayes, which is an actual Outlook plugin (it can be set up as a proxy-style of filter too, although I have not tried this yet) and set about training it.
SpamBayes record is 2 spams in the inbox in 750+ messages, or about 99.7%. I like it.
I know there are more filters out there, and I am sure there are some that are even more effective. However, it comes down to whether my time is better spent worrying about 0.3% of messages or just dealing with it. I think I'm done looking for now.
I used to use The Bat! for my email here in the office. It does a wonderful job of sorting mail and even has a Bayesian filter plugin included, called BayesIT! that supposedly does a good job of chopping down spam. That was not my experience at all. In fact, BayesIT! never once successfully filtered anything, even after training. I switched to Winkler's Bayes Filter Plugin and got significantly better results, with a success ratio around 95%. That's not so bad, but it still means about 50 spams in the inbox per thousand messages.
I have since been converted over to Outlook here in the office (required, due to the calendar and coordination tools) and again I found myself looking for a good filter. I have SpamEaterPro which is a heuristic filter, but it's accuracy has a lot to do with the diligence of maintaining the whitelists and blacklists and it has gotten squirrely in its stability. Outlook's built-in spam filter is ok, but it still is not as effective as what I had before. So, I turned it off and looked for a better way. After trying several filters, I found a freebie on SourceForge called SpamBayes, which is an actual Outlook plugin (it can be set up as a proxy-style of filter too, although I have not tried this yet) and set about training it.
SpamBayes record is 2 spams in the inbox in 750+ messages, or about 99.7%. I like it.
I know there are more filters out there, and I am sure there are some that are even more effective. However, it comes down to whether my time is better spent worrying about 0.3% of messages or just dealing with it. I think I'm done looking for now.
0 Comments:
Post a Comment
<< Home