An Explainable, Fairness-Audited Temporal Deep Learning Framework for At-Risk Learner Classification in Online Education

# Supplementary Material — Anonymised for Peer Review This repository contains the complete reproducibility package for the manuscript: **"An Explainable and Fairness-Audited Temporal Deep Learning Framework for At-Risk Learner Classification in Online Higher Education"** *Author information withheld for double-blind peer review.* --- ## Contents ### Notebooks - `Notebook1_DataAndTraining.ipynb` — Loads OULAD, builds temporal tensor (n, 4, 26), applies SMOTE, trains the proposed LSTM + three classical baselines (LR, RF, XGBoost), saves all model artifacts. Runtime: ~20 min on an NVIDIA Tesla T4 GPU. - `Notebook2_EvaluationFiguresStats.ipynb` — Loads trained models, computes SHAP (GradientExplainer), measures SHAP stability via 3×200 bootstrap, runs DeLong's test and McNemar's test, computes 95% bootstrap confidence intervals (1,000 resamples), performs the fairness audit (DPD, EOD, DIR), generates all 12 manuscript figures. Runtime: ~10 min on the same GPU. ### Figures - `fig2_pipeline_schema.svg` — Source vector file for Figure 2 (methodological pipeline overview). Editable; the PNG version shown in the manuscript was rendered from this SVG at 300 dpi. ### Predictions (reproducibility) All model predictions on the held-out test set (n = 5,820). Random seed = 42 throughout. - `predictions/y_true.npy` — Ground-truth binary labels - `predictions/y_scores_lstm.npy` — LSTM probability scores - `predictions/y_scores_lr.npy` — Logistic Regression probability scores - `predictions/y_scores_rf.npy` — Random Forest probability scores - `predictions/y_scores_xgb.npy` — XGBoost probability scores ### Results (tables) - `results/table2_with_CI.csv` — Table 2 with 95% bootstrap CI for 6 metrics × 4 models - `results/delong_results.csv` — Pairwise DeLong test results (AUC-ROC comparisons) - `results/mcnemar_results.csv` — Pairwise McNemar test results (classification-decision comparisons) - `results/fairness_results.csv` — Fairness audit: DPD, EOD, DIR for gender, IMD band, disability, age band --- ## Reproducibility All analyses use a fixed random seed (`SEED = 42`). With the correct dependencies (see below), running the notebooks in sequence will regenerate every numerical value, figure, and table reported in the manuscript. ### Dataset The Open University Learning Analytics Dataset (OULAD) is publicly available under a Creative Commons Attribution 4.0 licence (CC BY 4.0). It is not redistributed here; download it from the official source and place the seven CSV files in a directory of your choice, then set the `OULAD_PATH` variable in Notebook 1 Cell 2 accordingly. ### Dependencies ``` numpy < 2 pandas tensorflow (2.x) scikit-learn (1.3) shap (0.42) imbalanced-learn (0.11) xgboost statsmodels matplotlib, seaborn ``` Tested on Python 3.10 with Google Colab (T4 GPU). ### Execution ```bash # 1. Open Notebook 1 in Google Colab (or Jupyter) # 2. Adjust OULAD_PATH to point to the directory containing the 7 OULAD CSVs # 3. Run all cells (Runtime → Run all) # 4. When complete, 19 artifact files are saved to Drive # 5. Open Notebook 2 and run all cells # 6. Output: 12 figures (300 dpi), 4 CSV results files, 5 prediction arrays ``` --- ## Licence Supplementary material: Creative Commons Attribution 4.0 International (CC BY 4.0). ## Data availability OULAD: available from Kuzilek et al. (2017), Scientific Data 4:170171, DOI: [10.1038/sdata.2017.171](https://doi.org/10.1038/sdata.2017.171). All code and derived artifacts in this repository are released under CC BY 4.0 and can be freely reused with attribution upon de-anonymisation at article acceptance.