Understanding predictive models in education: enhancing transferability, explainability, and generalizability

This thesis explores the application of Data Science and Artificial Intelligence (AI) within the field of educational sciences, focusing specifically on leveraging machine learning (ML) techniques to enhance educational outcomes. The primary case study revolves around predicting the risk of low achievement among students using data from the national INVALSI test, which serves as a practical application to address common challenges in AI integration in education: transferability, explainability, and generalizability. In addition to low achievement, the thesis considers other educational issues, such as academic dropout and knowledge tracing, reflecting a broader perspective on student performance. A key aspect of this work is the incorporation of Informed Machine Learning (IML) principles, which facilitate the infusion of domain expertise and other knowledge sources into predictive modeling. This methodological and epistemological reflection underscores the importance of understanding how predictive models can be effectively employed in educational settings to inform policy and practice. Throughout the thesis, various strategies are proposed to tackle the identified challenges. For transferability, the potential for adapting models to different educational contexts is examined. The explainability of ML models is emphasized as essential for fostering trust among stakeholders and supporting informed decision-making processes. Additionally, generalizability is addressed through innovative approaches to student representation across diverse cohorts. By interconnecting these themes, this research aims to contribute to the understanding of predictive analytics in education and provides a framework intended to support the thoughtful implementation of AI solutions in educational settings, with the aspiration that it may lead to improved outcomes for students.