HGCR: A Heterogeneous Graph-Enhanced Interactive Course Recommendation Scheme for Online Learning

As one of fundamental tasks in online learning platform, Interactive Course Recommendation (ICR) aims to maximize the long-term learning efficiency of each student, through actively exploring and exploiting the student's feedbacks, and accordingly conducting personalized course recommendation. Recently, Deep Reinforcement Learning (DRL) has witnessed great application in ICR, which can gradually learn student's dynamic preference through multiple-round interactions, and meanwhile optimize the long-term benefit of student. However, when modeling the student's hidden and unknown interest, so-called latent interest, the existing DRL-based recommendation schemes didn't fully characterize and utilize the relationships among courses and other associated objects, such as teachers of courses and courses' concepts, which may hamper the system's learning of student's latent interest and lead to sub-optimal recommendation. To address the above issue, this paper proposes a novel DRL based personalized ICR scheme enhanced with the heterogeneous graph, HGCR, which smoothly combines the graph neural network with advanced deep Q-learning neural network. Specifically, the paper's contributions are threefold. First, the heterogeneous graph is explicitly built to characterize the relationships among courses, concepts, and teachers. Second, the course representation is formulated through graph attention network. Then, a student's latent interest is characterized with her/his interactive courses, which is then fed into the Double Dueling Deep Q-network (DDDQN) for interactive course recommendation. Finally, thorough experiments on two real educational datasets demonstrate the proposed framework outperforms the state-of-the-art DRL-based methods.