Fig 2. Accuracy of student models for different number of GCN layers in a teacher model.

Student_KD and Student_LSP represent the students trained by distilling the knowledge of classes, and knowledge of the embedded topological structure of a teacher, respectively. Student_Base corresponds to a model trained with the ground truth labels without the teacher. Note that our proposed MustaD (denoted as Student_MustaD) provides the highest accuracy in most cases. We also observe that MustaD provides much better performance for deep GCN with many layers, unlike competitors whose performances do not improve with more layers.