2
$\begingroup$

I need to detect rules in a data set that contains real-valued features. A simplified example of my samples, defined by features a, b, and c, and having class 0 or 1, might look like this:

2.34, 1, 5.46 => 0

1.23, 0, 7.81 => 1

For this data set, I am interested in detecting rules like:

a > 1 and a<3 and b==1 => class=0

I have checked out association rule algorithms in Weka and then searched online, and they do what I want, but they all seem to need binary features. My question: is there an algorithm that can do what I want, or do I need to hack my own on the basis of existing association rule algorithms?

$\endgroup$
5
$\begingroup$

It sounds like a decision tree algorithm (such as rpart in R) will do what you need. Most decision trees will find rules similar to the one you outlined above.

$\endgroup$
2
  • $\begingroup$ Thanks! I got so focused on association rules that I forgot completely about decision trees. $\endgroup$ – ACEG Mar 11 '14 at 15:53
  • $\begingroup$ 6 years later. Another guy, another user case but the same misfocusing. Just wanted to thank very much for pointing this out. $\endgroup$ – flashspys May 7 '20 at 20:57
1
$\begingroup$

You can normalize and discretize the feature ranges. If you have a training dataset with scores in continuous space (e.g. BM25 retrieval scores), it can be normalized and split into multiple categories such that [0.25-0.50], [0.51-075], [0.76-1.00]. Now the minimum support and confidence can be decided by the specific distribution.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.