Localized Actionable Explainability for MOOC Dropout Intervention Using Clickstream Learning Analytics

Massive Open Online Courses (MOOCs) have expanded access to learning, but learner dropout remains a persistent challenge in online education. Most dropout prediction studies transform learner traces into behavioral features and train machine learning or deep learning models to identify at-risk learners. Although this predictive layer is useful, it remains limited for educational decision-making because a risk score does not explain why a learner is at risk or what type of support should be proposed. This study introduces Localized Actionable Explainability (L-AXAI), a learning analytics framework that combines dropout prediction, local explanation, and personalized pedagogical nudging. The framework first transforms clickstream logs into weekly behavioral indicators, then trains predictive classifiers, and finally applies SHAP-based local explanation to identify the strongest risk driver for each high-risk learner. These local explanations are mapped to targeted pedagogical actions using a transparent rule-based translation matrix. Experiments on the KDD Cup MOOC dropout benchmark produced closely comparable results for the two strongest boosting models. Under the fixed stratified test split, Gradient Boosting achieved 85.79% accuracy, 86.57% ROC-AUC, and 94.41% PR-AUC, while XGBoost achieved 85.77% accuracy and the same ROC-AUC. Because repeated-seed evaluation and statistical significance testing were not conducted, the small difference between these point estimates should not be interpreted as definitive model superiority. The explainability analysis identifies model-associated patterns involving weekly activity intensity, active days, and early problem engagement. The intervention layer generates candidate nudge categories, mainly navigation, re-engagement, and session-completion suggestions. The framework therefore returns an estimated dropout risk, a local risk driver, and a candidate pedagogical action. The nudges are intended as low-intensity decision-support suggestions; their appropriateness, perceived pressure, and effectiveness require validation with instructors and learners.

@article{OUASSIL2026,
  title     = {Localized Actionable Explainability for MOOC Dropout Intervention Using Clickstream Learning Analytics},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Mohamed Amine OUASSIL and Mohammed JEBBEARI and Mouaad ERRAMI and Rabia Rachidi and Soufiane HAMIDA and Bouchaib CHERRADI and Abdelhadi RAIHANI},
  doi       = {10.14569/IJACSA.2026.0170734},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170734}
}