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

In the vast digital landscape of modern education, learning management systems record a constant stream of student activity: clicks on course pages, submissions of quizzes, participation in forums, and the timing of every login. For years, data scientists have used these digital traces to build models that predict which students are likely to struggle or fail. These predictive tools act like an early warning system, flagging at-risk learners before a crisis occurs. However, a prediction of trouble is only half the solution. Knowing a student is in danger does not tell a teacher or advisor what specific, practical steps can be taken to help them. A warning that says "this student will fail" is less useful than a guide that says "if this student reviews these three pages and attends one support session, their chances of passing improve significantly." The challenge lies in turning a statistical risk score into a concrete, actionable plan that respects the real-world limits of time, resources, and student privacy.

This is the central problem addressed by a new framework called SC2R, developed by researchers to bridge the gap between predicting failure and enabling success. The team, working with data from the Open University Learning Analytics Dataset, sought to move beyond simple risk scores to generate intervention plans that are not just mathematically possible, but actually feasible within the constraints of a real classroom. They recognized that a recommendation is useless if it asks a student to do something impossible, such as attending a class after the deadline has passed, changing a permanent personal characteristic, or accessing a resource that is not available. To solve this, they combined a predictive model with a system that checks every suggested action against a set of strict, machine-readable rules regarding timing, budget, and availability.

The researchers built a system that operates in three distinct stages. First, it looks at a student's current state, capturing their recent activity and past performance at specific points before an upcoming exam. A predictive model then estimates the probability of that student passing the next assessment. If the probability is too low, the system shifts to its second stage: generating a list of potential actions to improve the outcome. These are not vague suggestions but discrete, countable steps, such as reviewing a specific number of practice quizzes or revisiting a set of course pages. The system uses a mathematical optimization method to find the combination of actions that offers the best chance of success while requiring the least amount of effort from the student.

However, the true innovation of this work lies in the third stage, where the system acts as a rigorous gatekeeper. Before a plan is ever presented to a human, it is translated into a structured digital format and subjected to a semantic validation check. This process verifies that every proposed action fits within the reality of the educational environment. It ensures that the suggested activities can be completed before the exam deadline, that the total effort does not exceed a reasonable limit, and that the student is not asked to alter immutable facts about their identity or history. The system also checks that any required resources, like a specific lab session or support office, are actually open and available during the proposed time.

When the researchers tested this framework on data from over 127,000 student snapshots, the results highlighted the critical importance of these constraints. The predictive component of the system performed strongly, accurately identifying students at risk with a high degree of reliability. More importantly, the system successfully generated thousands of compact intervention plans, with the average plan requiring just over one action to achieve a desired improvement in success probability. The average cost of these interventions was calculated at roughly 10.3 units, a metric representing the effort required. Crucially, the semantic validation layer revealed that without these strict checks, many plans that looked good on paper would have been rejected in practice. In a controlled test of 200 cases, when the system was asked to consider the availability of resources, nearly 13 percent of the plans that initially seemed valid were flagged as impossible because the required resources were not available at the suggested times.

The study suggests that the value of such a system is not in guaranteeing that a student will pass, but in ensuring that the advice given is trustworthy and operationally sound. The researchers found that while the system could generate plans that were mathematically valid, the semantic constraints were necessary to filter out those that were practically unenforceable. This distinction is vital for building trust between educators and artificial intelligence. If a system suggests a plan that a teacher knows is impossible to execute, the entire tool loses credibility. By making the feasibility of a plan machine-checkable, the framework ensures that recommendations are not just model-valid, but also semantically feasible and interpretable.

The work also examined how stable these recommendations are when the underlying data changes slightly or when the model is retrained. The results showed that the plans were moderately stable against small changes in student input, but less stable when the predictive model itself was updated. This indicates that while the framework is robust for immediate decision-making, the recommendations may need to be revisited if the learning environment or the model's understanding of it shifts significantly. The researchers were careful to note that their evaluation was conducted offline using historical data, meaning they could not prove that following these plans would actually change student outcomes in a live setting. The study establishes the feasibility of the method, not the causal impact of the interventions.

Ultimately, this research offers a new way to think about educational support. It moves the conversation from simply identifying who is at risk to defining what can be done about it, provided that the solution respects the complex web of time, resources, and rules that govern education. By treating intervention planning as a problem that requires both mathematical optimization and semantic reality checks, the framework provides a path toward decision support that is as practical as it is predictive. The findings suggest that for artificial intelligence to be truly useful in education, it must be able to speak the language of the classroom, understanding not just the numbers, but the constraints that make those numbers meaningful.