Deep learning-based methods for evaluating and optimizing teaching quality in higher education

Abstract

The higher education sector continues to face significant challenges in ensuring consistent and objective teaching quality evaluation due to subjective assessment practices, non-standardized evaluation criteria, and limited real-time analytical feedback. Conventional statistical and machine learning approaches often experience feature redundancy, inter-class overlap, and insufficient temporal modeling, resulting in unstable and biased multi-class predictions. To overcome these limitations, this research proposes Big Bang Big Crunch optimized Peephole Long Short-Term Memory (BBBC-Peephole LSTM) intelligent teaching quality assessment. The Course Quality Evaluation Dataset (CQED), containing 10,000 records, is utilized, comprising structured instructional indicators categorized into three classes: Low, Medium and High. The methodology includes systematic preprocessing with missing value handling and Z-score normalization to ensure numerical stability and distribution consistency. Principal Component Analysis (PCA) is employed for feature extraction and dimensionality reduction, minimizing redundancy while preserving maximum variance. Within the proposed architecture, BBBC performs adaptive hyperparameter optimization, automatically tuning hidden units and learning rate to improve convergence efficiency and prevent suboptimal configurations. The Peephole LSTM is designed to capture long-term temporal dependencies in student engagement and instructional performance data through cell-state-aware gating mechanisms, thereby enhancing sequential learning precision. The model is implemented in Python using deep learning libraries for structured training and validation. Experimental results demonstrate an overall classification accuracy of 97.90%, confirming strong predictive performance, improved generalization capability, and balanced multi-class evaluation reliability in higher education systems.