Transformer-Enhanced Architecture with Adaptive Feature Selection for Performance Prediction in MOOCs

This study proposes a deep learning framework for predicting student performance in Massive Open Online Courses (MOOCs) by leveraging programming behavioral data, with a focus on comparing the effectiveness of Transformer-based architectures against traditional Convolutional Neural Networks (CNNs). To address the challenge of high-dimensional and noisy feature spaces in educational data, an adaptive feature selection strategy is introduced, integrating correlation analysis, SHAP value interpretation, and Recursive Feature Elimination (RFE) to prioritize critical predictors such as runtime error frequency and code complexity metrics. The Transformer model incorporates self-attention mechanisms and GRU-based temporal encoding to capture long-range dependencies in sequential programming activities, while the CNN serves as a baseline for local pattern extraction. Experimental results on a real-world MOOC dataset demonstrate that the Transformer architecture significantly outperforms CNN, achieving a 12.7 % improvement in prediction accuracy and superior robustness to sparse data. Ablation studies further validate the contribution of adaptive feature selection, reducing model training time by 34 % without sacrificing performance. This work provides a practical paradigm for educational data mining, highlighting the transformative potential of Transformer-based approaches in learning analytics.