Question on semi-supervised part
Closed this issue · 2 comments
benelot commented
Hello!
I am looking for graph neural network datasets with partially missing labels. As far as I understand from your paper, the WikiCS dataset can be used for node label prediction. Interpreting semi-supervised in this context, a network can use this dataset and learns from samples with and without labels. Am I right here? Assuming I am correct, I was trying to find such samples in the data, but could not find them easily, thus was a bit confused.
Thanks for your answer!
pmernyei commented
Hi! Yes, you are correct. In the context of semi-supervised node
classification, the entire graph is normally known at training but with
only a subset of the labels. The way the data is organized for WikiCS, all
node labels are given in a single vector, and there's separate binary
vectors specifying train/val/test masks. At training time, only the labels
allowed by the train mask are used.
By the way, note that this dataset (along with many others) have been
included in the torch_geometric GNN library, so you might find it
convenient to just load it from there:
https://pytorch-geometric.readthedocs.io/en/latest/modules/datasets.html#torch_geometric.datasets.WikiCS
Cheers,
Peter
…On Fri, Feb 26, 2021 at 2:59 PM Benjamin Ellenberger < ***@***.***> wrote:
Hello!
I am looking for graph neural network datasets with partially missing
labels. As far as I understand from your paper, the WikiCS dataset can be
used for node label prediction. Interpreting semi-supervised in this
context, a network can use this dataset and learns from samples with and
without labels. Am I right here? Assuming I am correct, I was trying to
find such samples in the data, but could not find them easily, thus was a
bit confused.
Thanks for your answer!
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#2>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AGBZPZDFYU324LL7XJRA2XTTA6SNZANCNFSM4YIPWAZA>
.
benelot commented
Oh ok I see! The training mask does not only mask samples as a whole, but masks also some of the labels only, thus describing a certain pattern of missingness. Thanks!