I found that note very interesting:
> consider "house" and "housing" ... the former is used less frequently in a spammy context then the later.
The point is that stemming is valuable unless P(spam|root) is substantially different from P(spam|derivedWord). But that data is available to you through Bayes. Has anyone tried a 'conditional stemming' system where the word is stemmed iff
|P(spam|root) - P(spam|derivedWord)| < someThreshold
? I wonder if that would improve classification accuracy. I suppose it might be a big performance hit on the stemming algorithm though.