Knowledge tracing (KT) captures the mastery status and proficiency level of students by modeling the answer information in their historical answering records, and then predicts their future answering situations. It can help teachers better grasp their learning situation and make more reasonable and targeted teaching plans. The previous methods mainly focus on the temporal changes in the learning process of students, ignoring the spatial relationships between students, exercises, and knowledge concepts. Graph is an effective way to model these relationships and several works have been shown it’s feasibility. However, existing graph-based KT methods mainly extract features from the predefined simple graphs, which usually is a single type of information in students’ answering records. To better utilize the abundant information hidden in multiple correlations, we propose a new model called Multi-Edge Features Enhancement for Graph-Based Knowledge Tracing (MEGKT). In MEGKT, we first model student historical learning records as two heterogeneous graphs, where there are three types of edges between students and exercises, students and students, exercises and exercises. Then, we explore the implicit information of student answering situations from the constructed heterogeneous graphs by employing Edge Aggregated Graph Attention Network (EGAT), to build different student models. On this basis, we integrate these student models into a stronger teacher model through online knowledge distillation, allowing the teacher model to supervise the learning of the student model to achieve better results and make predictions through the teacher model. To evaluate the performance of the proposed MEGKT, we do experiments on two real education datasets and the results show that our MEGKT has achieved state-of-the-art performance in predicting student answering exercises.