Artificial Intelligence (AI) is reshaping the education system around the world, and an important question remains open: what does an AI system model when it models a
learner? This thesis focuses on the progression around that question. It progresses from modeling an individual learner to modeling the quality of relationships and mutual understanding
between people involved in the learning program. The thesis is organized around two parts: Part I focuses on the knowledge tracing (KT) problem, and Part II
focuses on dropout prediction in apprenticeship programs that involve multiple stakeholders. 

The knowledge tracing task models or estimates a learner’s evolving knowledge state in order to predict their future performance. The initial exploratory KT study uses large language models (LLMs) for this task and identified two main findings. It identifies that LLM-derived skill relationship graphs contribute well to the graph-based KT model (GKT-LLM), whereas we also used LLMs directly as a sequence prediction model, which does not perform well (LLM-KT). These initial results motivate the HiSACKT approach, which is the main contribution of this part. HiSACKT is a hierarchical self-attention model with dynamic skill clustering to enhance KT. It combines attention across multiple layers or scales with a learnable skill clustering mechanism that enables the model to capture skill relationships directly from interaction data without relying on predefined skill graphs. The approach is evaluated on four publicly available datasets and achieves strong and consistent performance by reaching an AUC up to 93.33%, while also producing interpretable skill clusters.

Part II of the thesis focuses on apprenticeship programs that inherently involve multiple stakeholder-based settings. The main stakeholders of this setting include an apprentice, a workplace master, and a training provider. ESPM targets these multiple stakeholder settings by introducing ensemble stacking with a perception-aware meta-learning model. The model predicts dropout contracts by explicitly modeling stakeholder engagement asymmetry and perception gaps. It also contains a separate interpretability module that reveals a counter-intuitive finding, which highlights that engagement and employer features contribute more strongly to dropout predictions than performance-related features. These results motivate a discourse-analytic (ToM) study that focuses on the quality of mutual understanding by identifying the mentalization depth of master’s written feedback. It derives a typology of relational dysfunction and shows that the discourse quality features are statistically independent and complementary to behavioral engagement features. Finally, the Double Theory of Mind (DToM) framework is presented, which generalizes the relational perspective to teacher-facing AI systems by formalizing a three-layer cognitive structure, along with six design principles and research agendas. These contributions together advance personalized and relational learning analytics by tracing a progression from modeling individual knowledge states to modeling the mutual understanding that shapes the learning relationships.