Knowledge Tracing (KT) aims to model learners’ dynamic knowledge states and predict their future response performance. Most of the existing KT models overlook the importance of advanced exercise representations and other beneficial exercise information. Besides, the relationship between exercises and learners is neglected, and difficulty and discrimination differ from exercises to exercises. Therefore, we propose a Dual-mode Contrastive Learning-Enhanced Knowledge Tracing (DEKT) model, which can obtain fine exercise representations and generate better predictions. Specifically, our model incorporates three additional types of heterogeneous nodes: learner, difficulty, and discrimination to explicitly exploit the exercise information. Furthermore, we employ the dual-mode contrastive learning (path-based mode and schema-based mode) to strengthen and complement exercise representations with semantic and structural information. Finally, through the dual-temporal attention mechanism architecture (one for tracing learners’ knowledge states and the other for predicting), we acquire the long-term dependencies between exercises between time step 1 and t, as well as between time step 1 and \(t+1\) , helping us to capture learners’ knowledge states in a more effective manner.