Student performance prediction can support timely instructional intervention, but course-level engineering education datasets are often small, heterogeneous, and vulnerable to evaluation leakage. This study develops interpretable behavioral features for an Engineering Fluid Mechanics course and benchmarks traditional machine-learning, ensemble-learning, and deep tabular models under a leakage-safe evaluation protocol. Using a primary multi-cohort dataset of 400 students, a 5-fold outer cross-validation was conducted with all preprocessing steps, including median imputation and model-specific scaling, fitted only within the training folds through reproducible pipelines. Linear Regression achieved the best overall performance on the primary dataset (MAE=1.692±0.112;R²=0.841±0.021), while ensemble and deep tabular baselines did not provide consistent gains at this dataset scale. Pairwise tests on out-of-fold absolute errors with Holm correction further indicated that the performance advantage of Linear Regression was statistically significant against all competing models (all Holm-adjusted p<0.002). Feature analysis and ablation results show that engagement- and attendance-related indicators contain substantial predictive information, whereas the Learning Engagement Index mainly improves pedagogical interpretability rather than adding independent predictive signal. Supplementary validation on an independent dataset of 1500 students produced lower absolute accuracy (Linear Regression:MAE=3.953± 0.129; R²=0.734±0.015), suggesting domain shift and feature-schema differences. These findings highlight the importance of leakage-safe benchmarking, transparent baselines, and cautious interpretation when applying predictive analytics to small-to-medium engineering education datasets.
Wang et al. (Fri,) studied this question.