An Improved Differential Evolution Algorithm Using Mahalanobis Distance and Cholesky Decomposition for Wrapper-Based Feature Selection

In machine learning, feature selection is a fundamental component that plays a crucial role in improving prediction accuracy and reducing the computational time of classification models. Feature selection involves eliminating irrelevant or redundant features that can negatively affect model learning. This article proposes a wrapper-based method that uses a binary evolutionary algorithm for feature selection. The Differential Evolution algorithm performs feature selection and is combined with Mahalanobis distance and Cholesky decomposition (MaCRO-DE). Using Cholesky decomposition to calculate Mahalanobis distance helps to determine the proximity of feature subsets, which is useful for deciding which features to select or discard. Subsequently, two classifiers (K-nearest neighbors and the support vector machine) are used as evaluators, along with a V-shaped transfer function, to identify the most suitable option for the proposed method. To evaluate the effectiveness of the proposed methodology, the CEC-2017 benchmark and twelve widely used datasets from the UCI repository are employed. Experimental tests are conducted using thirteen prominent metaheuristic algorithms. The results indicate that the proposal effectively achieves its objectives, outperforming the benchmark across the performance indicators used for comparison, including Accuracy, F1-Score, Recall, Precision, Fitness value, and number of selected features.