Adaptive Quantum‐Inspired Optimisation of Neural Networks for Imbalanced Student Performance Prediction

Abstract

Student performance prediction (SPP) is a critical task in educational data mining (EDM) for facilitating proactive educational interventions. However, existing predictive models face specific challenges: the high pedagogical cost of misclassifying at‐risk students (class imbalance), the difficulty of feature selection in high‐dimensional noisy data, and the inherent variance of solitary stochastic optimisation. To address these limitations, this article proposes an Adaptive Quantum‐Inspired Optimisation Framework (SRA‐ANN‐E). The framework integrates three optimisation mechanisms: (1) a Cost‐Sensitive Quantum (CSQ‐SRA) mechanism that reconstructs the search topology via penalty barriers to prioritise minority class identification; (2) an Adaptive Collapse (AC‐SRA) mechanism utilizing a stagnation‐driven monitor to regulate the exploration‐exploitation trade‐off; and (3) an elite population ensemble strategy to mitigate stochastic variance. Validation was conducted across four heterogeneous datasets (OULAD, UCI, PISA 2022 and PISA 2015), encompassing structured logs, psychological attributes, and unstructured natural language processing features. When benchmarked against 11 baseline algorithms, SRA‐ANN‐E outperformed existing hybrid meta‐heuristics (e.g., TLBO‐ANN) and demonstrated a statistically significant, albeit marginal, advantage over modern tree‐based ensembles (e.g., XGBoost), achieving an 86.80% minority recall in dropout prediction. Furthermore, coupled with SHAP‐based interpretability analysis, the framework identified compact feature subsets, translating numerical outputs into interpretable pedagogical indicators (e.g., mathematics anxiety) while maintaining online inference efficiency. These results indicate the framework's viability as a reliable and interpretable decision‐support instrument for academic early warning systems.