Is it possible to partition Graph datasets?

Flower Datasets
1

Is its possible to partition graph datasets especially HeteroData with more than one node type and their edges?

I see that we can create a custom dataset and partitions from local data, but in this case the features and labels are assumed to be same across the samples.

Could you suggest and approach to go about with hetero graphs

2

Is it also possible to bypass FederatedDataset?

3

Welcome to the Flower community @boscojacinto!

Can you please elaborate what your FL setup is like? It sounds like you have a setup with multiple clients where some training data can be “shared” among clients?

Likewise, for your second question, can you expand a little what do you mean by bypassing?

4

Hi @chongshenng thanks for replying back.

Sure I can elaborate on my setup.

I have a global graph of Restaurants in a city. It is a heterogeneous graph with “Restaurants” as one node type, “Area” (in city) as another node type and “Customers” as another node type (Note: I may be adding another node type in the future).

I am using pyG’s HGTConv which operates over HeteroData type where the features of each node type are represented as dict of x(features) and y(labels) and the edges are represented as COO matrix.

My clients are going to be several restaurants which will train this global model, the restaurant will be having its Customer’s data on which it will train.

Firstly I want the server to partition the global graph as sub-graphs and send them to the clients (via context or messaging? ) Is it possible? Also is the partitioning only for the purpose of simulation or even for production?

Secondly, I did kinda by pass the FederatedDataset class in my PoC and managed to have FedAvg work on the results of atleast two client in the training rounds. I used the load_data function to load from local data.

I was unsure of this customization, but seems like the workflows are independent of the dataset structure, but it would be cool to have them compatible with Graphs especially Hetero graphs. A reference that I used is here
https://www.nature.com/articles/s41467-022-30714-9

Regards,
Bosco