Is it ok to generate a classifier by collecting all association rules so that the conclusion part refers to the target variable? does it perform better than decision tree or rule learning?
2
-
$\begingroup$ In order to have a classifier you have to have some input variables and some out variables. Which would be your input/output variables in this case? Which is the question that your classifier needs to answer? $\endgroup$– rapaioJan 14, 2014 at 16:55
-
$\begingroup$ @AurelianTutuianu There is no specific input/output variables or question for classifier. Just in general is it possible? $\endgroup$– SigaJan 14, 2014 at 20:21
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
There are several algorithms for creating classifiers out of association rules. The most basic approach will only add a default rule as the last rule, otherwise the resulting classifier may not be able to make prediction for some instances. More sophisticated algorithms such as CBA (Classification based on Associations) by Bing Liu also perform pruning, which selects only some of the discovered association rules.
Benchmark of CBA against C4.5 can be found in the original paper:
Ma, Bing Liu Wynne Hsu Yiming. "Integrating classification and association rule mining." Proceedings of the fourth international conference on knowledge discovery and data mining. 1998.
According to this evaluation (by CBA author) CBA has slightly better accuracy (1% difference) than C4.5.
A limited evaluation of the simplest approach that essentially just adds a default rule to the end of the rule set, is covered in:
Kliegr, Tomáš, et al. Learning business rules with association rule classifiers. Rules on the Web. From Theory to Applications. Springer International Publishing, 2014. 236-250.
This evaluation shows that using the list of association rules as is drops the accuracy compared to the prune baseline, typically by more than 1%. This indicates that C4.5 would perform better.
Disclaimer: I am co-author of the latter paper.
