<p>As an essential part of online education systems, knowledge tracing is crucial in predicting students’ mastery of each knowledge item and improving learning efficiency. Although recent knowledge tracing models have begun to exploit heterogeneous graph structures to capture complex relationships among students, exercises, and knowledge concepts, they often fail to effectively suppress noisy interactions and exploit higher-order semantic information. However, existing knowledge tracing (KT) models face problems such as noise interference, data sparsity, and a lack of higher-order semantic information when modeling learning behavior in complex educational scenarios. These issues limit models’ ability to accurately understand and predict the learning process. Therefore, this paper proposes the Trust-Aware Knowledge Tracing via Heterogeneous Graph Influence Propagation and Key Node Aggregation (TAKT) model to address these challenges. The model constructs a heterogeneous graph that integrates student interactions, accounts for direct influence and indirect communication contributions, and incorporates dynamic trust, while filtering out high-value nodes. In addition, TAKT bridges heterogeneous node semantics through semantic fusion across meta-paths, enabling multidimensional modeling of learners’ cognitive states. Extensive experiments on four public datasets demonstrate that TAKT consistently outperforms baseline models in terms of accuracy and AUC.</p>