Re: Interesting but incomplete
> % Interesting comparison.
> % I find, however, that not having
> % compared with Spamassassins built-in
> % Bayesian filter makes the comparison
> % incomplete. I use this myself, and
> % difference in results are striking.
> % spamassassin uses the results from
> % Bayesian filter not directly, but as
> % another factor weighing in on the
> % spamminess of an email, the results
> % really aren't directly comparable to
> % other Bayesian filters.
> % It would be excellent if you could
> % one or two of your tests with this as
> % well.
> I actually did that today, and
> surprisingly enough the performance
> isn't significantly different from
> spamassassin without the bayesian
> filter. In fact it scored slightly worse
> on the final "TCR" analysis - though as
> I said by an insignificant amount.
> Of course I may have done something
> wrong - I just deleted the contents of
> the $HOME/.spamassassin directory,
> touched $HOME/.spamassassin/user_prefs
> and then ran sa-learn on the training
> emails. And then ran spamassassin as
> My training sets are smaller than
> recommended by spamassassin, so it may
> be essentially ignoring them?
Spamassassin requires several hundred of both spam and ham messages before it will begin using the Bayesian filtering rules.