A Particle Swarm Optimization Based Model for Feature Selection

Analyzing and visualizing high-dimensional data with a large number of variables or features is a cost consuming and critical process because of the complex relationships among the features. The internal organization of datasets is often simplified and made more manageable for analysis using techniques such as dimensionality reduction and clustering. The proposed research introduces a particle swarm optimization based model for feature selection (PSOFS). This model is designed to enhance classification accuracy while simultaneously reducing the number of features in a dataset. The motivation behind this approach is the resource-intensive nature of feature exploration in large-dimensional datasets. PSOFS employs a two-phase process: initially selecting features using the Particle Swarm Optimization (PSO) algorithm, followed by an incremental application of the Laplacian Score (LP) to reduce the feature set and improve classification accuracy. The study evaluates the performance of PSOFS using three well-established classification methods: multi-layer perceptron, decision trees, and k-nearest neighbors. To benchmark PSOFS, seven existing feature selection algorithms are also employed. Results indicate that the proposed PSOFS approach consistently achieves higher average classification accuracy (91.36%) compared to the seven alternative algorithms (88.05%). Additionally, PSOFS achieves this superior performance while significantly reducing the average percentage of selected features.