Closed hmtbgc closed 2 months ago
Thanks for your excellent work. I am curious about experimental setting of SGFormer and your model, since the accuracy reported in this paper is lower than the original paper.
SGFormer achieves relatively high accuracy by randomly selecting 20 nodes per class for the training set. However, we adopt the public split provided by PyG, which is a more challenging setting widely used for comparing node classification performance across models. Interestingly, under the random selection of 20 nodes per class, our model can achieve over 86% node classification accuracy on the Cora dataset.
Thanks for your excellent work. I am curious about experimental setting of SGFormer and your model, since the accuracy reported in this paper is lower than the original paper.