Is this semi-supervised learning?

[RainZ11]

Hello, I'd like to ask a question. In many previous papers, the adjacency matrix of GCN includes all nodes, while in this paper, it seems to only include the nodes of the training set. So, isn't this semi-supervised training?

[RainZ11]

If unlabeled samples are not involved in the training process, wouldn't it mean that we cannot effectively utilize unlabeled samples, and would it have an impact on the results?

[RainZ11]

Consider an example: in the Cora dataset, with a compression rate of 0.1, the dimension of the adjacency matrix is 14. Does this represent 10% of the training nodes rather than 10% of the entire dataset? This truly troubles me.

[ChandlerBang]

In our experiments, we actually consider both transductive (semi-supervised training) and inductive settings.

in the Cora dataset, with a compression rate of 0.1, the dimension of the adjacency matrix is 14. Does this represent 10% of the training nodes rather than 10% of the entire dataset

It depends on how you define the compression rate. In Table 2 of our paper, the compression rate is defined as the ratio of condensed graph size (14 nodes) to the original graph size (2710 nodes), which would be 0.5%. Although in the code we have r to be 0.1, it is just for easy implementation and it is different from how we calculate the compression rate in pactice.

Hope this helps. Thanks.

[RainZ11]

Thanks very much!