In order to structure our dialogue, we'll divide data mining
into the following task,
each of these has its own tricks, tips, and traps.
- Generate: produce the data, somehow (you got to start somewhere);
- Classify: stamp each item in the data with some class label (e.g. what is the goal?);
- Statistics: summarize the current data (and that has
some interesting quirks)
- Discretize: clump the numerics into chunks (clotting
the experience base into denser units)
- Row
select/reject: anomlay detection, handle outliers, select relevant data, discard irrelevant
- Column select/reject: feature subset selection, prune superflous details,
- Train and test: finally- we run the learners!
- Evaluate: did we win anything?
- And, again: iterative learning
schemes.
Of course, there are many other ways to discuss data mining.
For example, some of the above distinctions are artifical
and some learners combine two (or more) of these tasks into one system.
But, for the sake of teaching, we'll step through things as above.
]]>