A maximum discrimination DEA method for ranking association rules in data mining

Abstract

A data mining algorithm, such as Apriori, discovers a huge number of association rules (ARs) and therefore efficiently ranking all these rules is an important issue. This paper suggests a data envelopment analysis (DEA) method for ranking the discovered ARs using a maximum discrimination between the interestingness criteria defined for all ARs. It is shown that the proposed DEA model has a unique optimal solution which can be computed efficiently when the maximum discrimination between the criteria, the difference between DEA weights, is considered. The contribution of this study can be explained as follows: First, we show that using the conventional DEA model for ranking ARs may produce an invalid result because the weights corresponding to interestingness criteria would not discriminate between the criteria. This is investigated for a dataset consisting of 46 ARs with four criteria, namely support, confidence, itemset value and cross-selling. The paper also introduces the maximum discrimination between the weights of the criteria and obtains the optimal solution of the corresponding DEA model efficiently without the need of solving the related mathematical models. On the other hand, this model concludes less number of useful rule(s). A comparative analysis is then used to show the advantage of the proposed DEA method.

Keywords:

• data mining
• association rule
• data envelopment analysis
• maximum discrimination
• ranking

2000 AMS Subject Classifications:

• 90B50
• 90C05
• 90C49
• 90C90
• 90C46

Acknowledgements

The authors are grateful to two anonymous reviewers and the editor of IJCM for their constructive comments that have improved the paper substantially. The authors also thank Dr Ali Emrouznejad at Aston University for his helpful comments on the revision of this article.