Semi-supervised node classification accuracy for Cora, Citeseer, and Pubmed.

We perform the distillation from trained teachers with various number of GCN layers: 2, 4, 6, 8, 16, 32, and 64. Student_MustaD is our distilled student that has the same hidden feature dimension as the teacher. Note that MustaD consistently outperforms other KD methods while preserving the multi-hop feature aggregation of the deep teacher.

我们针对具备不同图卷积神经网络（Graph Convolutional Network，GCN）层数的已训练教师模型实施知识蒸馏，教师模型的GCN层数分别为2、4、6、8、16、32与64。Student_MustaD为本研究中通过蒸馏得到的学生模型，其隐特征维度与对应教师模型完全一致。值得注意的是，MustaD在保留深度教师模型多跳特征聚合能力的前提下，性能始终优于其他知识蒸馏（Knowledge Distillation，KD）方法。