ACO Based Feature Subset Selection for Multiple k-Nearest Neighbor Classifiers

The k-nearest neighbor (k-NN) is one of the most popular algorithms used for classification in various fields of pattern recognition & data mining problems. In k-nearest neighbor classification, the result of a new instance query is classified based on the majority of k-nearest neighbors. Recently researchers have begun paying attention to combining a set of individual k-NN classifiers, each using a different subset of features, with the hope of improving the overall classification accuracy. In this paper we proposed Ant Colony Optimization (ACO) based feature subset selection for multiple k-nearest neighbor classifiers. The ACO is an iterative meta-heuristic search technique, which inspired by the foraging food behavior of real ant colonies. In ACO, real ants become artificial ants with the particular abilities such as distance determination & tour memory. The solution is constructed in a probabilistic way based on pheromone model in the form of numerical values. The concept of this approach is selecting the best possible subsets of feature from the original set with the help of ACO and combines the outputs from multiple k-NN classifiers. The experimental results show that this proposed method improves the average classification accuracy of k-NN classifier. KeywordsMachine Learning; k-Nearest Neighbor; Feature Subset Selection; Ant Colony Optimization.