C.M. Teng and R. Hewett (USA)
Traditionally association rules are generated according to two criteria: the minimum support and the minimum ac curacy (commonly called “confidence”) of the rule in the given data set. We examine the advantages and disadvantages of this approach, and propose an alternative formulation of association rules, based on confidence intervals, which makes explicit the role of inference in the mining of associations. In lieu of a point estimate, the rule accuracy is construed as falling within a certain interval with a certain level of confidence. We developed an algorithm for computing interval association rules, and by exploiting some observations about the relationship between the rules, some unnecessary processing can be avoided by selective pruning. Conceptually, interval association rules are grounded in the theory of statistical inference. We showed that while the standard framework is adequate in many situations, the interval association rule framework can give rise to more desirable results in a wider range of cases.
Important Links:
Go Back
