Re: DM: Re: problem of sample size

One useful way of analyzing minority classes is to use a rule miner which
accepts a consequent constraint (In this case, the consequent constraint
would be the minority class identifier).  This won't build a classifier for
you, but it will build a useful descriptive model (which can then be turned
into a predictive model with other techniques).

I have put the "dense-miner" algorithm on-line, which allows you to do just
this. If you can make your data-set HTTP accessible, you can try it out.

Here are a few things to note:

o This publically accessible version has several restrictions. For example,
it will only attempt data-sets of size 10 MBytes or less.

o Works only with categorical attributes (so continuous attributes must be
discretized).

Instructions are available at:

http://www.almaden.ibm.com/cs/people/bayardo/vinci/index.html

Roberto Bayardo

• Prev by Date: DM: I'm back...
• Next by Date: DM: Datasets for Mining
• Prev by thread: RE: DM: Datasets for Mining
• Next by thread: DM: I'm back...
• Index(es):
  • Date
  • Thread