Henry,
In the paper An Assessment of Case-Based Reasoning for Spam Filtering
http://www.comp.dit.ie/sjdelany/publications/AICS%202004%20(crc).pdf
the authors compare CBR and a naive Bayes (NB) with one conclusion (on their test data, with their implementation of NB) that daily updating of the training data using misclassified mails caused an improvement in FPs but a degradation in FN rate that led to an overall negative effect on their measure of performance.
How does that compare to your results on the effect of training and learn on error vs learn on everything?
If CBR does end up better than NB when used with learn on error, that is an advantage in terms of computational resources required.
-- sidney
