Predicting Student Academic Performance Using GenAI-Enhanced Deep Learning Features: A Comparative Study

Abstract

Predicting student academic performance is a critical challenge in educational data mining, with implications for early intervention, personalized learning, and policy-making. This paper evaluates whether Generative AI (GenAI)-derived behavioral and contextual features add predictive value beyond traditional educational attributes for grade classification. We benchmark five deep learning architectures — Residual MLP, Bidirectional LSTM (BiLSTM), CNN with Global Average Pooling (CNN-GAP), Feature Tokenizer Transformer (FT-Transformer), and a Hybrid CNN-LSTM model — together with four gradient-boosted and ensemble baselines (Random Forest, XGBoost, LightGBM, CatBoost), under two feature-set conditions: a Traditional baseline and a Full GenAI-augmented setting. Experiments are conducted on a curated dataset of 14,003 student records with 16 base features spanning study habits, technology use, psycho-social indicators, and learning style, expanded with GenAI-derived engagement, risk, technology-adoption, and support-need scores generated by an open-weight instruction-tuned language model (Qwen2.5-1.5B-Instruct). Contrary to our initial hypothesis, results do not support a general benefit from GenAI feature augmentation: across the nine models tested, only the FT-Transformer shows a statistically significant improvement (+3.3 percentage points, paired t-test, p = 0.0002), while Random Forest shows a significant decline (−0.6 points, p = 0.040); the remaining seven models show small, non-significant shifts in either direction (all p > 0.13). A controlled ablation study using the strongest architecture (Hybrid CNN-LSTM) reinforces this picture: removing the GenAI-derived features entirely (33.01% ± 1.32%) performs marginally better, not worse, than the full feature set (32.76% ± 0.89%), with overlapping error bars. The best absolute performance in the study is achieved by CatBoost on traditional features alone (34.6%), and the strongest overall configuration is a heterogeneous CNN-GAP + CatBoost ensemble (34.4% accuracy, mean AUC 0.665). We report these results transparently, including the negative and null findings, and argue that this architecture-dependent, largely non-significant pattern is itself the useful finding for practitioners considering GenAI-based feature engineering in educational data mining: it is not a reliable, drop-in source of predictive lift, and its value, if any, appears contingent on the specific model architecture used.