CBA (v2.0) -- Classification Based on Associations
CBA mines association rules and builds accurate classifiers using a
subset of association rules. See the paper in KDD-98, which can also
be downloaded from http://www.comp.nus.edu.sg/~liub.
CBA v2.0 is the second version of CBA. This version has the following
new features:
1) Mining association rules with multiple minimum supports (SIGKDD'99)
2) Better classification accuracy
-- After testing against the 26 datasets used in our KDD-98 paper (the
datasets are all from UCI repository for Machine Learning), CBA v2.0
achieves the average error rate of 15.2% over these datasets. For
these data sets, C4.5 (release 8) achieves the error rate of 16.7%
(c4.5rules) and 17.3% (c4.5tree), while for CBA v1.0, it is 15.8%.
See the detailed comparision in:
http://www.comp.nus.edu.sg/~dm2/result.html
-- The better result is obtained due to the use of multiple minimum
class supports (rules with different classes have different
automatically assigned minimum supports). The multiple minimum
class supports scheme is a special case of multiple minimum
supports above.
3) Faster rule mining
4) A HTML viewer to help user analyze the discovered rules
CBA v2.0 can be downloaded from: http://www.comp.nus.edu.sg/~dm2
Furthermore, we have also made a post-analysis system downloadable. The system
is called Interestingness Analysis System (IAS). IAS helps the user find
interesting (conforming and unexpected) rules based on his/her existing
domain knowledge.
Regards
Bing Liu
Email: liub@comp.nus.edu.sg
Web: http://www.comp.nus.edu.sg/~liub
Copyright © 1999 KDnuggets