Abstract
In many real world applications, comprehensibility of the classifier is as important as its accuracy. The medical field is one of those where this requirement is more pronounced. It is not enough for users in this field to have an accurate classifier, and they also need to verify and analyze the logic of the classification process. It is difficult to have confidence in a black box type of classifier when the classification decision is a matter of life and death of a patient. In recent years, algorithms for classification rule discovery based on the ant colony optimization meta-heuristic (ACO,) have been proposed, which fulfill both the requirements of high accuracy and comprehensibility. This paper reports some improvements in a recently proposed ACO based classification algorithm, called CAntMiner, whose main feature is a heuristic function based on the compatibility of pairs of attribute-values and class labels, and its application on medical datasets. We study the performance of the algorithm for twelve commonly used datasets and compare it with ten well known classification algorithms, three of which are ACO based. Experimental results show that the accuracy rate obtained by CAntMiner is better than that of the compared algorithms. We also discuss some other issues related to comprehensibility of the classifier building process.