Extensions to the ant-miner classification rule discovery algorithm

Abstract

Ant-Miner is an application of ACO in data mining. It has been introduced by Parpinelli et al. in 2002 as an ant-based algorithm for the discovery of classification rules. Ant-Miner has proved to be a very promising technique for classification rules discovery. Ant-Miner generates a fewer number of rules, fewer terms per each rule and performs competitively in terms of efficiency compared to the C4.5 algorithm (see experimental results in [20]). Hence, it has been a focus area of research and a lot of modification has been done to it in order to increase its quality in terms of classification accuracy and output rules comprehensibility (reducing the size of the rule set). The thesis proposes five extensions to Ant-Miner. 1) The thesis proposes the use of a logical negation operator in the antecedents of constructed rules, so the terms in the rule antecedents could be in the form of <attribute NOT= value>. This tends to generate rules with higher coverage and reduce the size of the generated rule set. 2) The thesis proposes the use stubborn ants, an ACO-variation in which an ant is allowed to take into consideration its own personal past history. Stubborn ants tend to generate rules with higher classification accuracy in fewer trials per iteration. 3) The thesis proposes the use multiple types of pheromone; one for each permitted rule class, i.e. an ant would first select the rule class and then deposit the corresponding type of pheromone. The multi-pheromone system improves the quality of the output in terms of classification accuracy as well as it comprehensibility. 4) Along with the multi-pheromone system, the thesis proposes a new pheromone update strategy, called quality contrast intensifier. Such a strategy rewards rules with high confidence by depositing more pheromone and penalizes rules with low confidence by removing pheromone. 5) The thesis proposes that each ant to have its own value of α and β parameters, which in a sense means that each ant has its own individual personality. In order to verify the efficiency of these modifications, several cross-validation experiments have been applied on each of eight datasets used in the experiment. Average output results have been recorded, and a test of statistical significance has been applied to indicate improvement significance. Empirical results show improvements in the algorithm's performance in terms of the simplicity of the generated rule set, the number of trials, and the predictive accuracy.