Feature Selection in Classification using Binary Max-Min Ant System with Differential Evolution
Jennifer Joyce M. Montemayor, Renato V. Crisostomo · 2019
Ant Colony Optimization (ACO), an algorithm based on the natural foraging behavior of ants, has been used as a feature selection method to maximize the performance of the classifier or minimize the number of features in various classification problem applications. The multi-agent nature of the algorithm and indirect coordination mechanism make it attractive for searching a large feature space for an optimal feature subset. However, the iterative nature of the algorithm has a tendency to stagnate and unable to converge toward optimum solutions. This study presents a feature selection method based on the algorithmic variant of ACO called Max-Min Ant System combined with Differential Evolution (BMMASDEFS). Differential Evolution (DE) is introduced to the pheromone update mechanism of Max-Min Ant System (MMAS) to influence the search towards optimal solutions. The performance of the proposed algorithm is tested on well-known benchmark datasets from the UCI Machine Learning Repository and is compared with the results of similar feature selection methods. The proposed algorithm generated feature subsets that are 30% to 60% smaller than the original feature set. The experimental results show that the generated feature subsets improved the performance of the classifier than when all the available features are used. It is also observed that the performance of the classifier using the generated feature subsets is competitive with the results obtained using similar feature selection techniques.