next up previous
Next: Comparison of the qg Up: Expert-Guided Subgroup Discovery: Methodlogy Previous: The Expert's Role in

4. Analysis of the Proposed Rule Quality Measure Used in Heuristic Search

It is well known from the ROC analysis, that in order to achieve the best results, the discovered rules should be as close as possible to the top-left corner of the ROC space. This means that in the TPr/FPr tradeoff, TPr should be as large as possible, and FPr as small as possible. Similarly, in the TP/FP space, TP should be as large as possible, and FP as small as possible.

In this work the quality measure qg = TP/(FP+g) using generalization parameter g has been defined. This section explains why this quality measure has been selected, in comparison with other more intuitive quality measure like a cost-based measure qc involving `cost' parameter c.