Knowledge Tracing (KT) is an essential task of an intelligent tutoring system. It aims to predict students’ performance in answering questions by modeling their knowledge mastery. It supports personalized learning for students based on their historical learning records. In practice, online learning platforms provide numerous questions, but students typically engage with only a fraction of them. This makes it difficult for KT models to accurately predict answers to other questions. This problem can generally be alleviated by mining the relationships between questions. However, the relationships have been proven to be intricate and complex. Most of the existing KT methods concerning relationship mining fail to comprehensively discover the connections. To address this issue, we propose IRKT, integrating relationships from internal features and external manifestations for knowledge tracing. We first use heterogeneous hypergraph networks (HHN) to learn high-order relationships between questions from their internal features. Then we excavate the relationships from external manifestations by measuring correlation. Correspondingly, graph convolutional networks (GCN) are applied to learn the pairwise relationships. To optimize question representations, we employ contrastive learning (CL) to integrate two types of relationship information. Experiments demonstrate that our proposed method outperforms previous classical methods in terms of AUC on three widely used datasets.