Re: [Ifile-discuss] Effect of widely differing volumes on ifile classification

[Jason Rennie]

address@hidden said:
> Recently, the rate, which had been consistent for some time, began to
> plunge to about 50% and stayed there, until I deleted .idata and
> rebuilt it from scratch, and it's now classifying better than before.
> (Data attached at bottom for completeness) 

Are there any discernible differences between your current collection of 
e-mail (what you used to rebuild .idata from scratch) and the collection 
used to build the old .idata?  Do you keep all of your e-mails?  Can you 
tell us about the types of misclassifications?  Did it look pretty random 
or were there certain folders that ifile seemed to send everything to?

address@hidden said:
> This didn't happen - it actually started to misclassify the mailing
> lists which receive all the volume.

That's strange, the mailing lists should be very easy for it to correctly 
classify.

FYI, the anomalous behavior may just be a relic of the fact that ifile
uses Naive Bayes to do classification.  I could go into details, but they
may be more confusing than illuminating...  Naive Bayes can do some weird
things when the training data is highly skewed.

Jason