SC2R made student-risk recommendations machine-checkable without claiming causal improvement

Ngoc Luyen Le, Marie-Hélène Abel and Bertrand Laforge address a gap in student-risk analytics. A predictive model may identify a learner as likely to struggle, but a probability does not tell an adviser what can reasonably change. Generic counterfactual explanations can suggest altering variables that are immutable, unavailable, too late, or outside a learner's control. The authors propose SC2R, a semantics-constrained counterfactual recourse framework designed to make proposed intervention plans operationally meaningful.

SC2R combines four components. A calibrated predictive model estimates risk. Integer programming searches over discrete action variables for compact changes that could alter the prediction. A lightweight RDF vocabulary represents the intervention plan. SHACL validation checks constraints such as timing, budget, immutability, and service availability. This separation is important: mathematical optimization can find a model-valid plan, while semantic validation asks whether the plan is permitted and feasible in the actual educational setting.

The framework is evaluated offline on the Open University Learning Analytics Dataset using snapshots before assessments at two decision horizons. The authors report strong predictive performance, scalable generation of compact plans, and cases where semantic validation exposes infeasible recommendations that an optimization-only approach would accept. Their claim is carefully bounded. They do not say that following a generated plan causes a student to improve. They show that recourse becomes more actionable when recommendations can be checked against explicit operational constraints.

Several limitations remain. OULAD is historical and institution-specific. Recorded clicks, assessments, and demographics do not capture every reason a learner may disengage or every support available. A plan that changes a model prediction may not change learning, wellbeing, or persistence. Integer constraints encode institutional judgments that can be incomplete or inequitable. Calibration and feasibility can drift as courses, services, populations, and policies change. Students and advisers were not shown testing the recommendations in a live decision process.

A responsible pilot would begin with low-stakes, supportive actions and participatory review. Advisers, teachers, disability services, learners, and data-governance staff should define which variables are actionable and which must never become targets. Every recommendation should show its evidence, constraints, cost, timing, uncertainty, and alternative options. A human adviser should be able to reject it, while a student should be able to question the data and decline an intervention without penalty. Prospective evaluation should measure service access, burden, fairness, unintended effects, and actual outcomes.

For AIEDHK, SC2R illustrates a mature distinction between prediction and decision support. The technical contribution is not merely a better risk score; it is a representation that lets people and machines test whether a proposed response fits real rules. The explicit refusal to claim causality is equally valuable. Educational analytics should be judged by whether recommendations are feasible, contestable, equitable, and beneficial in practice, not by whether an optimization can move a probability across a threshold. Publishing rejected plans and the constraints that rejected them would also help institutions audit whose circumstances the formal rules fail to represent.