Why the shape of data.y could be arbitrary?

[evantianzg]

Hi everyone,

I noticed that the description of data.y is as follows:

"data.y: Target to train against (may have arbitrary shape), e.g., node-level targets of shape [num_nodes, *] or graph-level targets of shape [1, *]"

I am a little confused about the above description and got two questions:

We all know that a label is an 1D vector. So the shape of data.y targets node-level should be [num_nodes,]. But why the shape of data.y is [1,]? Why isn't it the number of graphs? i.e. [num_graphs]?

Another question is why the shape of data.y could be arbitrary?

[EdisonLeeeee]

why the shape of data.y is [1,]? Why isn't it the number of graphs? i.e. [num_graphs]?

That's because a graph is represented as a data object, where each data (graph) is associated with a scalar label y represented as a one-dimensional vector [1, *].

why the shape of data.y could be arbitrary?

This is an intentional design to provide more flexibility to accommodate different scenarios.
In certain cases, nodes may have multiple labels, such as in the PPI dataset. In such cases, the labels are represented as 2D vector.

[evantianzg]

Hi Edison,

Thanks for your reply. It makes sense and is really a good design.