A Modified Ant Colony Optimization with KNN for High-Dimensional Data Classification

High-dimensional datasets are common for various real-world problems. Though these datasets contain useful information, they are hard to classify using a machine learning algorithm due to the curse of dimensionality. Feature selection is a significant machine learning concept that aims to generate optimal feature subsets from a high-dimensional feature space. In this project, we proposed a modified ant colony optimization (ACO) feature selection algorithm incorporating two new rules. The first rule modified the standard heuristic information gain measurement, while the second modified the pheromone update. The resultant feature subset generated by this modified ACO is fed into a k-nearest neighbor (KNN), and the resulting algorithm is called KACO. The performance of KACO was evaluated on five benchmark datasets, and the results were compared with KNN and KNN with sequential backward elimination (KNN&SBE) algorithms. The results show that KACO outperformed KNN, and KACO outperformed KNN&SBE in four datasets. The results also show that the global search method used in KACO can locate optimal feature subsets in high-dimensional feature space.