Thema: Can Foundation Models Replace Tradit
ional Machine Learning in Educational Data Mining?
A Case-Study on Sma
ll Educational Datasets
Beschreibung: Educationa l Data Mining (EDM) applications frequently rely on relatively small tabula r datasets collected from individual courses, study programs, or institutio ns. Such datasets often contain only a few hundred or a few thousand instan ces, limiting the effectiveness of data-hungry machine learning approaches and increasing the risk of overfitting. Recent advances in foundation model s for tabular data have introduced a new paradigm in which models are pre-t rained on millions of synthetic learning tasks and can subsequently perform classificationand regression on previously unseen datasets with little or no task-specific training. Despite their promising results in various domai ns, their applicability to educational datamining remains largely unexplore d. This research investigates whether foundation models can serve as a viab le alternative to traditional machine learning algorithms in educational pr ediction tasks involving limited amounts of data. We conduct an empirical e valuation of TabPFN (Prior-Data Fitted Networks), a foundation model for ta bular data, against widely used baseline methods, including Random Forest, XGBoost, CatBoost, Support Vector Machines, and Multi-Layer Perceptrons. Ex periments are performed on freely available educational datasets covering s tudent performance prediction, academic success forecasting, and dropout ri sk detection. In addition to predictive performance, we analyze robustness to small sample sizes, sensitivity to class imbalance, and model stability across repeated validation procedures. The results indicate that foundation models achieve competitive and often superior predictive performance on sm all educational datasets while requiring substantially less hyperparameter tuning than conventional approaches. Furthermore, the performance advantage becomes particularly pronounced in low-data scenarios where traditional mo dels struggle to generalize. The findings suggest that foundation models ma y represent a promising new direction for educational data mining, especial ly for institutions with limited historical data.
LOCATION:Raum Z 312 DTSTAMP:20261109T124359 DTSTART:20261110T153000 DTEND:20261110T163000 END:VEVENT END:VCALENDAR