Interestingness via what is not interesting
Sigal Sahar · 1999
As the size of mined databases increases, the sheer volume of rules produced by data-mining algorithms often overwhelms their users.These users only want to be presented with a short list of interesting rules.Some work has been conductedon determining what is interesting, with an agreement that interestingness is ultimately subjective.While there is no single, formal definition of interestingness, there is a consensus that finding interesting rules is a very difficult problem, requiring domain knowledge and/or user interaction.Our objective is to take the first step to a comprehensive solution to this problem in any domain.We present a simple and short process of eliminating a substantial portion of uninteresting association rules in a list outputted by a data-mining algorithm.Instead of trying to establish what is interesting, we look for rules that are not interesting; more specifically, the simple rules whose elimination implies the automatic elimination of many other rules in the list.Thus, our process requires very low-intensity user interaction: in three iterations, a user usually eliminates about 30% of the rules.A little over five iterations will usually halve the size of the problem.This elimination is a significant first step since size is the essence of the problem.We present representative results of executions of the algorithm over three real databases: WWW logs, grocery store transactions and adult census data. 1