Inquiry is a foundational skill for lifelong learning, critical thinking, and democratic participation. Open-ended learning environments (OELEs) offer a powerful way to foster such skills through authentic, self-directed exploration. Yet for these tools to be truly impactful, it is not always sufficient for learners to simply engage with them as they are often complex to navigate; educators and researchers must also be able to make sense of learner behaviour, determine when and how to intervene, and design support that accommodates the varied ways students approach inquiry.

This thesis aims to address key limitations in the modelling of inquiry learning by centring diversity, generalisability, and interpretability as core design principles. It responds to five persistent challenges in the field: the poor scalability of existing models across tasks and domains; the difficulty of modelling in low-data, high-variance contexts; the predominance of post-hoc analyses that preclude timely support; the tendency to overlook learner diversity in behaviour and background; and the limited integration between computational modelling and educational insight. Across these challenges, the thesis develops methods that are theoretically grounded, empirically robust, and practically applicable, advancing both the technical modelling of inquiry and the human contexts in which such models are deployed.

We begin with a generalisable framework for predicting students' conceptual understanding early in the learning process using only their interaction logs. Evaluated across two different user studies, our models achieve high predictive accuracy after just a few actions while maintaining interpretability. We show how inquiry behaviours such as experimental setups relate to different levels of conceptual mastery, enabling real-time support.

We then explore how inquiry strategies transfer across simulations and domains. Using clustering and representation learning, we find that learners tend to maintain consistent inquiry profiles across tasks. However, when strategy shifts do occur, especially when transitioning from simpler to more complex simulations, they are often linked to improved learning outcomes. This insight opens the door to generalisable modelling of inquiry as a transferable cognitive skill.

This thesis also addresses algorithmic biases head-on by investigating how demographic differences, shaped by broader societal structures, can lead to biased algorithmic outcomes. We develop a diagnostic framework to quantify how these behavioural differences affect model performance and show that signal variation, not just data imbalance, can drive unfairness. In response, we propose mitigation strategies based not on demographic attributes, but on behavioural profiles. These approaches improve model fairness without sacrificing accuracy, and generalise to multiple learning paradigms including simulations and educational games.

This thesis places a strong emphasis on real-world applicability and empirical validation. Across 12 user studies involving hundreds of students in Switzerland, the U.S., Canada, and Colombia, we evaluate the technical contributions under conditions of data scarcity and curricular diversity.

By combining algorithmic design with cross-context behavioural analysis and fairness-aware modelling, this work lays the foundation for scalable, and fair timely educational AI systems.