Abstract
Accurate prediction of student academic performance plays a vital role in enabling early intervention, adaptive instruction, and data-driven educational decision-making. However, the effectiveness of deep learning models such as Long Short-Term Memory (LSTM) networks is often constrained by suboptimal hyper-parameter tuning and insufficient exploration of the search space. To overcome these limitations, this paper introduces a novel hybrid learning framework, CNAE-LSTM, which integrates Chaotic initialization, a Niche-based evolutionary mechanism, and Alpha Evolution (AE) into a unified optimization paradigm termed Chaotic Niche Alpha Evolution (CNAE). Specifically, a chaotic operator is first employed to enhance population diversity during initialization; then a niche-based grouping strategy partitions the population into three subgroups with distinct mutation probabilities to preserve diversity and avoid premature convergence; finally, an Alpha Evolution mechanism guides the evolutionary search toward promising regions with improved exploration–exploitation balance. The proposed CNAE is utilized to optimize key LSTM hyper-parameters, including network architecture and training settings. Extensive experiments conducted on a real-world secondary education dataset demonstrate that CNAE-LSTM consistently outperforms conventional LSTM, SVM, CNN, Transformer models, as well as LSTM enhanced by QGA, QPSO, CWOA, QGWO, and grid/random search strategies. Results show that CNAE-LSTM achieves superior predictive accuracy, stability, and generalization capability, validating the effectiveness of the proposed hybrid framework for improving academic performance prediction.