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

Details

Learning analytics models can flag at-risk students but typically produce recommendations educators cannot act on — suggestions that violate timing windows, budget constraints, or resource availability. SC2R addresses this by adding semantic feasibility validation on top of counterfactual recourse generation.\n\nThe pipeline combines a calibrated predictive model with integer-programming-based recourse over discrete action variables. An RDF vocabulary represents intervention plans, and SHACL validation enforces operational constraints: timing, budget, immutability, and resource availability. Evaluated on the OULAD dataset at two decision horizons, SHACL validation reveals infeasible plans that lighter optimization-only methods would accept without question.\n\nThe paper carefully avoids overclaiming: it does not assert that generated plans improve student outcomes causally. This epistemic honesty is appropriate and rare — the gap between 'model says this should help' and 'this intervention is proven effective' remains large. The contribution is making recommendations operationally feasible and machine-checkable, not proving they work.\n\nFor edtech practitioners, the framework offers a pragmatic approach to closing the last mile between a risk model and an action an educator can actually take. The semantic validation layer is an independent, reusable component that could be applied to other recommendation systems in constrained operational environments.