Repository navigation
Question about Predicted Label Dimensions and the Usages of Test Node Label #3
Description
Thanks for your great work, I have a quick question regarding label dimension and Figure 3 in your paper. In the inductive attention mechanism, the t×c dimension—does the c-dimension correspond to the label dimension of the training set or the test set? Additionally, how does the Linear GNN trained on the training set ensure that the output matches the label dimension of the test set if the label dimension is not consistent, such as training on cora and inference on arxiv? I understand the label-permutation invariance of the distance of LinearGNN predictions, but how can the linearGNN predict the exact test label dimension without any labelled test data? Does GraphAny need to use some test node labels to get the W for linearGNN?
Thanks in advance.
Activity
[-]Question about Inductive Predicted Label Attention Dimensions[/-][+]Question about Predicted Label Dimensions[/+][-]Question about Predicted Label Dimensions[/-][+]Question about Predicted Label Dimensions and the Usages of Test Node Label[/+]DuanhaoranCC commentedon Mar 14, 2025
Hello, I have similar questions regarding author paper.
As I understand, the linear GNN mentioned in the paper is essentially SGC, Laplacian transformations, and other message-passing schemes. However, I could not find the analytical solution explicitly stated. Could you please point out where the analytical solution is discussed in the paper?
Additionally, entropy normalization and distance normalization seem to serve the purpose of stabilizing and smoothing the model.
Finally, as you mentioned, the GraphAny model trained on Cora has only 7 classes, while the Arxiv dataset has 40 classes. How are the results on new datasets reported in the table obtained? Are they based on few-shot learning, in-context learning, or zero-shot evaluation?
Thank you in advance for your time and clarification!
AndyJZhao commentedon Aug 10, 2026
Hi all! Thanks for your interest. For your questions on "How to generalise to inconsistent label dimensions":
Generally, GraphAny uses an inductive attention module to learn the map of txt -> tx1 and uses the tx1 inductive attention fuses txc predictions. Let's say we train Cora with 7 classes (c=7), and generalize to Arxiv with 40 classes:
As you may notice, the cool thing is, c and c' does not appear in our leaned inductive attention module (txt -> tx1). That is, as long as we keep t types of LinearGNNs (e.g. LinearSGC1, LinearSGC2, Linear, LinearHGC1, LinearHGC2) fixed, we can transfer to any graph.
I hope this explains a bit how GraphAny learns. Further reading: Michael Galkin's blog on Graph/Geometric Foundation Model
Sorry for the late reply. I didn't received notices of issues.