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

Learning analytics models often identify students at risk of poor performance, but standard risk scores do not tell instructors or advisors which specific interventions are feasible and actionable. Existing algorithmic recourse methods typically operate as geometric perturbations in abstract feature space, which can generate mathematically valid recommendations that violate real-world educational constraints like deadlines, budgets, and resource availability. This paper asks whether counterfactual recourse in education can be made more operationally meaningful by combining predictive modeling, integer-programming-based recourse generation, and semantic validation.

The paper introduces SC2R (Semantics-Constrained Counterfactual Recourse), a framework consisting of three primary components. First, a calibrated logistic regression model estimates a student's probability of success on upcoming assessments at specific time-to-deadline horizons (such as 14 or 7 days prior). Second, an integer programming solver generates compact counterfactual intervention plans over discrete action variables, aiming to minimize intervention burden while crossing a target success threshold. Third, a semantic validation layer uses a lightweight RDF vocabulary alongside SHACL (Shapes Constraint Language) constraints to check whether candidate plans satisfy operational rules.

The semantic layer enforces explicit, machine-checkable conditions that pure optimization models often miss. These include timing constraints ensuring all actions occur before assessment deadlines, budget constraints bounding total intervention effort, immutability constraints preventing changes to fixed learner traits, and availability constraints matching resource schedules. Evaluated offline on the Open University Learning Analytics Dataset (OULAD), the results demonstrate that SC2R successfully generates compact intervention plans at scale, while SHACL validation identifies and rejects infeasible recommendations that lighter optimization-only approaches would otherwise accept.