An optimal feature selection method using a modified wrapper-based ant colony optimisation

In feature selection, applications often require very high-dimensional data. Feature selection algorithms are therefore designed to identify the relevant feature subset from the original features, which can facilitate subsequent analysis such as classification and clustering. Also, feature reduction helps to reduce dataset dimensionality, lessen the running time, and/or improve the prediction accuracy. In this paper, a new wrapper-based feature selection approach is proposed based on ant colony optimisation (ACO). In the proposed approach, an ACO search environment is built and every ant probabilistically selects attributes depending on the pheromone and heuristic values linked with every edge. Furthermore, a heuristic function is used along with the values of pheromone for the selection of the ideal attribute subset. Naïve Bayes classifier is used to compute the fitness of each selected feature subset. The computed classification accuracy from naïve Bayes classifier is used as a fitness function. Different datasets are used for the experimental evaluation of the proposed approach. The experimental results of the proposed technique are very promising. The proposed technique increased accuracy by 5 % on average during experimentation on all datasets used when the subset feature selection is performed. Moreover, in 9 out of 15 datasets, the accuracy is improved when the feature subsets are selected using the proposed technique and the existing genetic search technique.