In intelligent educational tutoring systems, knowledge tracing (KT) is essential for predicting students’ future performance based on their previous exercise sequences. Although deep learning methods are widely applied in knowledge tracing models, data sparsity remains an issue because many questions are linked to only a few skills and are answered by few students. Previous graph embedding models fail to fully capture question representations due to the sparsity of the question-skill graph. This paper proposes a graph-based KT model that incorporates question interactions to improve the quality of question representations. By extending the question-relationship graph with student response interactions, we alleviate graph sparsity. We employ a weighted graph convolutional network (GCN) to learn robust question representations and use a question-interaction attention layer to capture interaction information of questions and integrate corresponding skill relationships. Experiments on four public datasets demonstrate that the proposed model improves by 5.25% compared to conventional methods.