KDAriane

h1. CMAR - Classification with multiple association rules

This Ariane graph implements CMAR[1], a well known supervised classification method with association rules.

!https://forge.greyc.fr/attachments/195/cmar1.png!

The parameters are the following:

1. the number of validations

2. the absolute minsup threshold

3. the number of accepted exceptions

The input is the well know _iris_ database, which has 3 classes (on the first column).

The cross-validation is handled by the _for_ loop, which will accumulate the result of the process. That is why there is a _touch_ input, which creates an empty file with the _echo -n_ command.

* the XML may be viewed by a browser

!https://forge.greyc.fr/attachments/196/cmar2.png!

In the validation loop:

* the database is first split in training and testing set with the _repartition_ operator

* comments (lines starting with a sharp) are removed

* the number of classes is computed with the purple macro, which cut the first columns, sorts it and counts

* the classification decision is computed in the for loop on each class and is accumulated in the touched input

* the classification result is a set of columns, one per class, containing the vote value for each instance. It is pasted to the train set.

* a score operator computes various indicator in an XML file: recall, precision, score, confusion matrix, area under the ROC curve (ineffective when more than two classes)

* the XML is accumulated

!https://forge.greyc.fr/attachments/197/cmar3.png!

The model is computed as follows:

* _mvminer_ computed the non redundant rules (a free or minimal or generator patterns as antecedent, its closure as consequent)

* the rules whose antecedent are not minimal for a given item are removed

* the rules concluding on the focused class are kept

* they are measured by a Chi square

* the cover of the rules over the training set are kept

* the rules vote for the class

* the result is accumulated.

h3. References

[1] Li W., Han, J. and Pei, J. (2001). CMAR: Accurate and Efficient Classification Based on Multiple Class-Association Rules. Proc ICDM 2001, pp369-376.