Advice Needed: Transitioning from table data regression models to GNN

[heaynking]

Hi,

I'm currently working on transitioning from using regression models with table data to exploring Graph Neural Networks (GNNs).
This is my first time working with graph networks, and I would greatly appreciate any advice or insights you can provide.

For example, suppose we have the following four tables.

ID	Name	Location	Phone	Target
0	L.H	A3	B2	0.8
1	H.R	A2	B1	0.7
…	…	…	…	…
N	T.S	A4	B6	1.0

ID	Feature1	Feature2
L.H	54	43
H.R	156	36
…	…	…
T.S	14	26

ID	A-Feature1	A-Feature2	A-Feature3
A1	3	1	177
A2	6	5	244
A3	4	5	453
A4	19	8	132

ID	B-Feature1	B-Feature2
B1	400	36
B2	300	46
…	…	…
B3	145	51

There are more columns and rows in the data, but I simplified it to illustrate.
I want to use the relationships between these to create graph data and perform a task to predict the target for new data.
The new data will be a combination of the items listed in the above tables.

---

First, I would like advice on what kind of graph data and tasks should be set.
Currently, I'm considering the following two:

1. It is a task of making predictions at the graph level.

2. The other is the task of making predictions at the node level.

As shown in the table, the number of features is different, so that's a point of concern that should be noted.

I know the model's performance depends on the data's quality or the architecture itself, but I would appreciate it if we could discuss which approach is more effective in such cases.

[rusty1s]

What do you mean by the number of features is different? As long as they are consistent on the table-level, you should be absolutely fine. This is a perfect use-case for PyG, and we have tried to write many resources on it:

• https://pytorch-geometric.readthedocs.io/en/latest/tutorial/load_csv.html
• https://arxiv.org/abs/2312.04615

[heaynking]

Thank you very much for the answer!!
I mean that the number of node features is different. So, It should be heterogeneous data.

What do you think about the regression task?
Is it ok to use a graph-level prediction task? Or should I use a node-level prediction task instead?
Are there any guidelines or good ways to determine this?

[rusty1s]

From the example, it looks like this is a node-level prediction task since a target is attached to a specific node. Let me know if I am misunderstanding.