Current seminar · 2026
Every Semester Is a Distribution Shift: Trustworthy AI for Personalized Programming Education
How early and how reliably can students who need support be identified, and what does it take to act on such predictions responsibly?
Read the abstract
Introductory programming courses are a well-known point of early attrition, yet signs of difficulty usually reach instructors only after major assessments, when little time remains to intervene. This talk examines how early and how reliably students who need support can be identified, and what it takes to act on such predictions responsibly, drawing on a project at the University of Doha for Science and Technology (UDST), funded by the Qatar Research, Development and Innovation Council (QRDI), on AI-driven personalized learning and gamification in programming education. Using data from 650 students in 36 class sections across two academic terms, the team developed six regression models to forecast final-exam performance from information available one month before the exam. Under nested cross-validation within a term, the best model exceeded 91% accuracy; when models trained on one term were applied to the next, the best accuracy fell to approximately 84% and the ranking of the models changed. A comparison of the two terms’ correlation structures suggests why: assessment grades became stronger predictors of the final exam, while completion-time features shifted inconsistently, some weakening and others reversing in sign. The talk argues that, in educational settings, distribution shift is a structural property rather than an exceptional event: each term introduces a new cohort, revised assessments and new tools, and an early-warning model that is not actively maintained degrades precisely where its errors are most consequential. The talk also addresses student modelling for personalization, including the inference of learning styles from uneven MOOC interaction logs without questionnaires and the characterization of students through player typology, personality, learning style and achievement emotions, together with the limits of such inferred labels. It concludes with the requirements that educational AI places on trustworthy machine learning: drift-aware evaluation and maintenance, predictions that instructors can interrogate, safeguards against unfair labelling of students, and privacy-preserving deployment.