Classification-Rule Discovery with an Ant Colony Algorithm

Ant colony optimization (ACO) is a relatively new computational intelligence paradigm inspired by the behaviour of natural ants (Bonabeau, Dorigo & Theraulaz, 1999). The natural behaviour of ants that we are interested in is the following. Ants often find the shortest path between a food source and the nest of the colony without using visual information. In order to exchange information about which path should be followed, ants communicate with each other by means of a chemical substance called pheromone. As ants move, a certain amount of pheromone is dropped on the ground, creating a pheromone trail. The more ants follow a given trail, the more attractive that trail becomes to be followed by other ants. This process involves a loop of positive feedback, in which the probability that an ant chooses a path is proportional to the number of ants that have already passed by that path. Hence, individual ants, following very simple rules, interact to produce an intelligent behaviour – a solution to a complex problem – at the higher level of the ant colony. In other words, intelligence is an emergent phenomenon; that is, “the whole is more than the sum of the parts”. In this article we present an overview of Ant-Miner, an ACO algorithm for discovering classification rules in data mining (Parpinelli, Lopes & Freitas, 2002a, 2002b). In essence, in the classification task each case (record) of the data being mined consists of two parts: a goal attribute, whose value is to be predicted, and a set of predictor attributes. The aim is to predict the value of the goal attribute for a case, given the values of the predictor attributes for that case. To the best of our knowledge, the use of ACO algorithms (Bonabeau, Dorigo & Theraulaz, 1999; Dorigo et al., 2002) for discovering classification rules is a very under-explored research area. There are other ant algorithms developed for the data mining task of clustering – see for example Monmarché (1999) – but that task is very different from the classification task addressed in this article. Note that Ant-Miner was designed specifically for discovering classification rules, rather than for solving other kinds of data mining tasks. In other research areas ACO algorithms have been shown to produce effective solutions to difficult realworld problems. A detailed review about many other ACO algorithms (designed to solve many other different kinds of problems) and a discussion about their performance can be found in Bonabeau, Dorigo and Theraulaz (1999) and Dorigo et al. (2002). A typical example of application of ACO is network traffic routing, where artificial ants deposit “virtual pheromone” (information) at the network nodes. In essence, the amount of pheromone deposited at each node is inversely proportional to the congestion of traffic in that node. This reinforces paths through uncongested areas. Both British Telecom and France Telecom have explored this application of ACO in telephone networks.