From Student Risk Prediction to SC2R: Semantics-Constrained Counterfactual Recourse for Educational Decision Support

Synthesis: Le, Abel, and Laforge (2026) tackle a persistent gap in Learning Analytics: prediction models can identify students at risk, but they rarely say which interventions are feasible, actionable, and compatible with educational constraints. They introduce SC2R, a semantics-constrained counterfactual recourse framework that couples a calibrated predictive model with integer-programming-based recourse generation over discrete action variables, a lightweight RDF vocabulary for representing intervention plans, and SHACL validation to enforce timing, budget, immutability, and availability constraints. Evaluated offline on the OULAD dataset, SC2R shows that compact intervention plans can be generated at scale and, crucially, that semantic validation catches infeasible plans that lighter optimization-only approaches would accept. Rather than claiming causal improvement, the paper demonstrates that counterfactual recourse becomes more operationally meaningful in education when recommendations are not only model-valid but also semantically feasible and machine-checkable.

Beyond prediction: the recourse gap

Learning analytics has made real progress in predicting failure, disengagement, and dropout from educational traces, assessment records, and learner profiles. But predictive performance alone is insufficient for decision support: instructors, advisors, and student-support services need recommendations that are actionable and operationally feasible, not just risk scores. This is the actionability gap — the gap between "this student is at risk" and "here is what can actually be done about it, within real constraints."

The paper positions counterfactual explanations and algorithmic recourse as the bridge between prediction and action. Recourse asks what should change to obtain a more desirable outcome, with an emphasis on actionability for the affected individual. But in education, naive recourse is often mathematically valid yet unenactable: methods that operate in feature space can recommend interventions that violate timing, budget, or availability realities (e.g., suggesting a session that happens after the assessment, or an action the institution cannot actually deliver).

The SC2R framework

SC2R assembles four components into one pipeline:

1. A calibrated predictive model for identifying at-risk students, providing reliable probability estimates.
2. Integer-programming-based recourse generation over discrete action variables — the framework searches for compact sets of feasible actions rather than arbitrary feature-space perturbations.
3. A lightweight RDF vocabulary for representing intervention plans in a structured, shareable, machine-readable form.
4. SHACL validation to enforce semantic constraints — timing (before/after assessments), budget, immutability (can't change a completed grade), and availability (whether an intervention can actually be delivered).

The framework is evaluated offline on the OULAD dataset, using snapshots constructed relative to each assessment at two decision horizons.

Findings

• Strong predictive component: the calibrated model provides reliable at-risk identification as a foundation for recourse.
• Scalable plan generation: compact intervention plans can be generated at scale across the dataset.
• Semantic validation matters: SHACL validation reveals infeasible plans that lighter optimization-only settings would otherwise accept — the paper's central methodological claim. Semantic feasibility and machine-checkability make recommendations genuinely operational rather than merely model-valid.
• Honest about causality: the authors deliberately do not claim causal improvement in student outcomes. The contribution is demonstrating that recourse becomes more operationally meaningful when it is semantically constrained.

What this means for practice

• Designers. Encode timing, budget, immutability, and availability as machine-checkable constraints before any plan reaches an advisor: the paper's central result is that optimization alone accepts plans which violate them, and the SHACL layer was what caught the infeasible cases.
• Designers. Use a calibrated logistic regression as the reference predictor (AUC 0.884 at the d−14 snapshot and 0.889 at d−7) rather than reaching automatically for nonlinear models — XGBoost and the TabTransformer-style model reached only 0.898 on the same d−14 split, a modest gain over the calibrated baseline.
• Designers. Generate recourse over discrete action variables rather than feature-space perturbations, so each recommendation is a set of deliverable actions the institution can actually schedule and check.
• Administrators. Present validated plans as decision support for instructors, advisors, and learning-support staff rather than as automated decisions; the framework's stated value is structuring and validating candidate plans before any practical use.
• Researchers. Test the framework on richer pedagogical interaction data and report robustness across retrained models, since the current action vocabulary is simplified and retrain stability was the weaker robustness indicator.

Limitations

• The evaluation is offline, model-based, and observational on the OULAD dataset; the authors state it does not establish that following a generated plan would improve student outcomes.
• The action vocabulary is simplified and derived from structured learner activity signals rather than richer pedagogical interaction data, which narrows the interventions the framework can express.
• Retrain stability is the weaker indicator, meaning some recommendations may remain sensitive to changes in the learned predictive boundary.
• The counterfactual baseline (a Wachter-style search) was compared on a controlled 200-case subset only, while the full-scale integer-programming plans were scored offline (127,972 plans).

Citation

Le, N. L., Abel, M.-H., & Laforge, B. (2026). From Student Risk Prediction to SC2R: Semantics-Constrained Counterfactual Recourse for Educational Decision Support. [cs.IR]. https://doi.org/10.48550/arXiv.2608.17618