Questions about building graphs with pos

[lfangyu09]

Hi,

Thank you so much for building this library. It really helps me a lot.

May I ask a few questions related to building graphs:

1. When I tried to use the absolute coordinates in data.pos, the results were bad. I found in Issue 3326, you mentioned we could use absolute coordinates in data.pos for some GNN layers. Otherwise, you suggested using edge_attr to consider the relative distance. Do you have any suggestions about processing the absolute coordinates?
2. What is the meaning of data.pos ? Is it only a key in a dictionary for normal GNN? Only some models like XConv could explicitly use data.pos ?
3. For the node feature, I have a feature that is continuous. It does not change across different nodes but changes across different time steps. Do you have any suggestions about processing this feature?
4. For the edge feature, I have a feature that is similar to the node type. It does not change across different time steps but does change across different nodes. Do you have any suggestions about processing this feature? If I want to set this feature as the edge type, how can I do it?

Thank you so much for your reply! I really appreciate your help!

[rusty1s]

1. Both ways can be found in literature. In general, using absolute coordinates no longer makes your GNN layer translation invariant, and thus might not generalize well to positions never seen before during training. If positions are equally distributed over training and test sets, using absolute coordinates should work quite well as well though.
2. Yes, it is just a separate attribute to differentiate node features from node positions. Following (1), you can also do x = torch.cat([data.x, data.pos], dim=-1) and interpret them as node features.
3. You should be able to just use it as a regular feature.
4. You can encode it as a one-hot vector if it models a type/index. Alternatively, there exist GNN layers that can take edge_type information into account, e.g., RGCNConv.

[rusty1s]

Usually, the Embedding approach is better since it is more memory and computation friendly. You avoid to store the one-hot matrix and avoid computing the embeddings via matrix multiplication.

[lfangyu09]

Thank you so much for your helpful reply.

I have a question about data.pos. I want to do the node regression, while I want to explicitly use data.pos. I checked some models, such as PointTransformerConv, but they did not include edge_attr. May I ask which model could explicitly consider data.x, data.pos, and data.edge_ettr? Thanks.

[rusty1s]

How do you want to use data.pos? I suggest to either include it into x (as a feature to encode absolute positions) or edge_attr (as a feature to encode relative positions between nodes).

[lfangyu09]

I used to use GNS model from the paper (Learning to Simulate Complex Physics with Graph Networks). It did (1) data.x= torch.cat((data.x, data.pos), dim=-1) and (2) use relative positions between nodes in edge_attr.
I would like to try other models, such as PointTransformerConv. However, these models actually include data.pos into x or edge_attr. Is it correct? Thanks.

[rusty1s]

Yes, this will happen internally.