A Fibonacci-Based Dynamic Neural Network Architecture for Imbalanced Classification of Higher Education Performance

Abstract

This study aims to address two problems in neural network design: the lack of a principled method for choosing hidden-layer neuron counts, and the reliability loss caused by severe class imbalance. To meet this objective, the study propose the Fibonacci Dynamic Neural Network, contributing, to our knowledge, the first architecture unifying number-theoretic hidden-layer sizing with imbalanced-learning correction, an unbiased selection property proven formally, and cross-domain validation. Conceptually, candidate hidden-layer widths are generated from the Fibonacci sequence: because consecutive terms grow proportionally toward the Golden Ratio, the resulting set spans a non-redundant, well-distributed range of capacities, unlike arbitrary sets from conventional methods. One candidate is drawn per layer via a discrete uniform distribution, giving every option equal selection probability, while class imbalance is corrected through synthetic minority oversampling applied only to training data. The architecture was evaluated experimentally on three benchmark datasets spanning higher education, credit card fraud, and charitable donation prediction, with imbalance ratios from about eighteen-to-one to nearly five hundred eighty-to-one, tested across five seeds against five baselines and an ablation study via the Wilcoxon signed-rank test. It achieved macro F1-scores of approximately 0.83, 0.94, and 0.81, improvements of sixteen, five, and sixteen percent over the strongest baseline, with the smallest training-test performance gap of any model tested; oversampling and Fibonacci spacing contributed most to these gains. These findings suggest number-theoretic structures offer a practical, reproducible basis for architecture design under class imbalance, useful for future work on automated range selection and for deploying reliable classifiers in imbalanced real-world settings.