Towards Transparent and Interpretable Predictions of Student Performance Using Explainable AI

Artificial Intelligence (AI) is increasingly being adopted in educational contexts to support data-driven decision-making, particularly in predicting student outcomes. However, the opaque nature of many high-performing models raises concerns around fairness, accountability, and interpretability which are factors that are especially critical in high-stakes environments such as GCSE examinations. This study investigates how Explainable AI (XAI) techniques can enhance the transparency and interpretability of machine learning models used to predict GCSE English Language and Mathematics performance. Using a real-world dataset from a secondary school in England, this research developed and evaluated predictive models, including Histogram-based Gradient Boosting (HGB) and a Multi-Layer Perceptron (MLP), to estimate student achievement outcomes. To address imbalances and maximise performance, the pipeline incorporated data pre-processing, feature engineering, and fairness-aware resampling strategies. The final HGB model achieved strong predictive accuracy while maintaining robustness across subgroups. To ensure interpretability, four XAI techniques namely SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), PDP (Partial Dependence Plots), and ALE (Accumulated Local Effects) were applied. These methods provided insight into the most influential features driving predictions, including attendance, CAT3 scores, SEN status, and EAL. Novel explainability metrics such as transparency score, explainability ratio, and interpretability ratio were proposed to systematically evaluate explanation quality and model clarity. In addition to technical evaluation, the study employed a stakeholder-centred design to assess how teachers, school leaders, and students interact with and interpret model explanations. Mixed-methods user studies revealed that personalised, context-sensitive explanations improved stakeholders’ decision confidence, supported intervention planning, and prompted critical reflection. Concerns were also raised about fairness, overreliance, and the ethical implications of demographic profiling. The research demonstrates that explainable models can enhance trust, transparency, and pedagogical utility when appropriately designed and evaluated in real-world educational settings. By integrating technical rigour with ethical and user-centred evaluation, this work contributes to the development of responsible, interpretable AI systems that align with the values and needs of educators and learners. The study offers both methodological innovations and practical recommendations for the responsible deployment of XAI in education.