Mining association rules with single and multi-objective grammar guided ant programming

Abstract

This paper treats the first approximation to the extraction of association rules by employing ant programming, a technique that has recently reported very promising results in mining classification rules. In particular, two different algorithms are presented, both guided by a context-free grammar that defines the search space, specifically suited to association rule mining. The first proposal follows a single-objective approach in which a novel fitness function is used to evaluate the individuals mined. In contrast, the second algorithm considers individual evaluation from a Pareto-based point of view, measuring the confidence and support of the rules mined and assigning them a ranking fitness. Both algorithms are verified over 16 varied data sets, comparing their results to other association rule mining algorithms from several paradigms such as exhaustive search, genetic algorithms, and genetic programming. The results obtained are very promising, and they indicate that ant programming is a good technique for the association task of data mining, lacking of the drawbacks that exhaustive methods present.