How Can AI Explain School Success?

Imagine having a magic wand that tells you which habits will make you shine at school. Scientists are trying to create it using artificial intelligence (AI). In a recent study, they used a machine‑learning model called LightGBM to analyze synthetic data about imagined students. The trick? They added a technique called SHAP, which reveals which factors push the model to make certain predictions. The results show that study hours, class attendance, and parental involvement are the "big players" for forecasting exam scores. Although the data are synthetic—made up to test the method—the experiment proves that a transparent analysis pipeline is possible, letting teachers and administrators understand the "why" behind predictions. This matters because many traditional algorithms are black boxes: they give an outcome but not the reasoning. With tools like SHAP, schools can move from "the model says I’ll fail" to "here’s what I need to improve." The authors stress that, for now, the findings don’t apply to real life, but they open the door to future research with real data, where AI could become a partner for personalising learning and supporting students more effectively.

The adopted approach combines LightGBM, a gradient‑boosting algorithm based on decision trees, with SHAP, which relies on Shapley game theory to assign each feature a marginal contribution value. LightGBM is renowned for efficiency on large datasets and its ability to capture non‑linear interactions among variables. However, its ensemble nature makes direct interpretation of decisions difficult. SHAP resolves this by computing, for every observation, the average marginal value of a feature across all possible insertion order permutations, ensuring additivity and consistency. In the educational context, synthetic features include continuous variables (study hours, participation score) and categorical ones (school type, socioeconomic level). Synthetic data were generated via a multivariate probabilistic model, preserving realistic correlations (e.g., positive correlation between attendance and grades). SHAP analysis highlighted that the most influential features have global importance (mean absolute SHAP) above 0.15, while minor variables contribute negligibly (<0.02). Moreover, SHAP dependence plots revealed non‑linear interactions, such as a threshold effect for study hours: beyond 15 weekly hours, the impact on grades spikes dramatically. These findings suggest that, once validated on real data, XAI models could inform policy decisions, for instance by directing resources toward parental involvement programs or remedial interventions for low‑attendance students.