Optimizing Multi-Layer Perceptron for Predicting Learner Migration Patterns: A Methodological Exploration

This research presents a novel approach to multilayer perceptron's (MLP) hyper-parameter optimization in solving learner migration problems in Limpopo, South Africa. While acknowledging the presence of various hyper-parameter optimization techniques, their applicability, strengths, and limitations differ. Our approach utilizes meta-heuristics, offering an efficient and adaptable method for complex search spaces and global exploration of optimal solution candidates. The social ski-driver (SSD) algorithm -originally designed for optimizing support vector machines (SVMs)- and cultural algorithm (CA) were utilized to determine the optimal hyper-parameter configuration for the MLP. The MLP was intended to predict the likelihood of a learner migrating, the reasons for migration, and the distance the learner will migrate to the next school. The two optimizers were run on sample data split into five folds, producing ten hyper-parameter sets (five pairs). The MLP was then built with each parameter set and subsequently run on a new dataset partitioned into five folds. The model performance results were compared using evaluation metrics such as the f1 score, variance analysis, and the Wilcoxon Signed-Rank test. There were no significant performance differences between the SSD and CA hyper-parameters, demonstrating the effectiveness of the SSD algorithm in optimizing neural networks. The CA-derived parameter set was selected due to its lowest variances across the datasets and its strong alignment with the convergence principles of the Berger-Tal multidisciplinary framework on the exploration-exploitation trade-off, providing a solid foundation for our findings.