Graph Attention Network - why using single graph

I am new to Graph Network and just start to practice this example provided by Matlab. Here is the link https://www.mathworks.com/help/deeplearning/ug/multilabel-graph-classification-using-graph-attention-networks.html.

Does anybody know why they combined all of the graphs into one signle graph(in their Preprocess Mini-Batch Function) in the model? Why not feeding in the graphs separately(each graph has 23 nodes) into the model? Thanks in advance for answering

Here is the part they mentioned about the one single graph.

Answer 1

[Dhruv]

Hi,

I understand that you are referring to the MathWorks documentation example where multiple graphs have been combined into a single graph for Graph Attention Networks (GATs).

The following reasons highlight the advantages for the same:

• Efficiency: Training and deploying a single GAT model on a combined graph is more efficient. This is because the GAT model can share parameters across all the graphs, which reduces the memory and computational requirements.

• Performance: Combining multiple graphs into a single graph can also improve the performance of GAT models. The GAT model can learn to aggregate information from all the graphs, which can help it to better understand the relationships between nodes in the graphs.

• Flexibility: Consolidating graphs also provides flexibility for training GAT models on diverse graph types. This approach simplifies handling varying graph sizes and complexities within a dataset.

In the example mentioned in the documentation, the graphs are all relatively small and have the same number of nodes (23 nodes). This makes it easy to combine the graphs and train a single GAT model on the combined graph. It is also possible to train GAT models on individual graphs. However, this is typically done when the graphs are very large or have different numbers of nodes.

I hope the solution provided above helps with your query.

[ch_tsai]

Thank you!