Bridging accuracy and interpretability in predictive learning analytics: A human-centered grey-box approach

Learning analytics has been increasingly using predictive models to support educational decision-making. However, the trade-off between predictive accuracy and interpretability remains a central challenge that limits how these models can be understood and used by educators. This study examines grey-box approaches to performance prediction by comparing two strategies: a theory-driven approach that groups interaction data into pedagogically meaningful categories, and a data-driven approach based on raw clickstream events followed by post hoc eXplainable Artificial Intelligence (XAI) techniques. Using clickstream data from three educational contexts with different scales and instructional formats, we empirically compare the predictive performance, interpretability, and pedagogical relevance of both strategies. The results show that raw-event models achieve higher predictive performance in the large-scale online context with rich interaction data, where XAI methods provide relatively stable and consistent explanatory patterns. In the smaller-scale blended setting, where platform data capture only part of the learning activity, theory-driven aggregation provides a predictive advantage, although results become inconclusive in very small cohorts. The findings indicate that there is no single optimal approach, and that modeling strategies should be selected according to data characteristics, instructional design, and analytic goals. From a human-centered learning analytics perspective, the study highlights the importance of methodological alignment for the meaningful and responsible use of predictive models in education.