Tailoring educational support with graph neural networks and explainable AI: Insights into online learners' metacognitive abilities

Abstract

Metacognition—the awareness and regulation of one's thinking processes—plays a crucial role in self-regulated learning (SRL), yet traditional educational research methods struggle to capture how metacognitive abilities manifest in actual learning behaviors. While computer-assisted learning (CAL) environments offer rich opportunities to observe these processes in action, educational researchers have typically analyzed this data using approaches that fail to connect metacognitive abilities with the complex, sequential nature of SRL behaviors. Our study bridges this gap by examining how 49 university students' metacognitive abilities shaped their learning patterns over one semester. We introduced a novel methodological approach that transforms diverse digital traces into unified graph structures, allowing us to map connections between metacognitive abilities and the planning, monitoring, and evaluation phases of SRL. Using attributed graphs, we integrated both static indicators and sequential behavioral patterns to predict metacognitive abilities with significantly higher accuracy than traditional single-data approaches, including Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Artificial Neural Networks (ANN), and Random Forest (RF). Through Explainable AI techniques, we revealed that high-metacognitive learners exhibited comprehension-centered, goal-oriented strategies across learning phases, while low-metacognitive learners focused primarily on task completion with limited strategic planning. These insights enabled us to develop personalized metacognitive profiles that can guide targeted educational interventions. Our approach demonstrates how advanced analytical methods can transform educational data into meaningful insights about cognitive processes, offering educators new ways to understand and support students' metacognitive development.