Artificial intelligence and machine learning models in education have shown exceptional performance and promise more accessible and personalized education. However, actual applications of these models remain rare as the best-performing models are usually the least explainable. Opaque models not only contain the risk of algorithms or automated decision-making systems making decisions that unfairly disadvantage certain groups of students, but they also prevent educational stakeholders from understanding the decisions. Thus, explainability plays a pivotal role in ensuring the right features are used and for detecting algorithmic discrimination. The THRIVE project aims to address the explainability issue by jointly considering (i) the representation and abstraction of data, (ii) the identification of “right” features with causality, (iii) the architecture of educational models, and (iv) model-usefulness established by the educational domain experts.

2023

2027