Graph sampling conditional on node/edge attr

[ssharpe42]

Are there any existing implementations or suggestions for node samplers where you can sample time directed subgraphs.

For example sample neighbors based on edge/node values so that all k hop neighbors v of node u have t_v < t_u

[rusty1s]

Such functionality is sadly not supported yet, but it would be quite important to add. I'm not entirely sure how the concrete interface should look like though. Shall t be given as a static vector during initialization of the sampler, or might it be important to apply such masking of nodes dynamically?

[ssharpe42]

Im envisioning this as a way to allow for training static GNN models with a single snapshot of a graph, but only allowing relevant information to be available for each node's neighborhood. Since t is fixed in a static graph (nodes/edges appear at time t_i) what do you mean by masking nodes dynamically?

Do you mean for example masking 2nd hop neighbors based on the directly sampled neighbors? I think there are two possibilities here:

1. A strict time directed graph sampling where each sampled neighborhood forms a directed graph that points towards the target node
2. A graph where all edges and nodes occur before the target node. This differs from 1 where a 1st hop neighbor occurs at t=2, 2nd hop neighbor occurs at t=4 and target node occurs at t=6. Not a time directed graph, but all multihop neighbors of the target node are present at t=6.

[rusty1s]

I see. I personally think that the second option makes more sense. In that case, sampling would have an interface similar to:

TemporalNeighborLoader(data, ..., temporal_attr='node_timestamp')

where temporal_attr points to the attribute in data that holds temporal information (either node-level or edge-level vector). I think the TemporalNeighborLoader would need to be different than our default NeighborLoader, as we need distinct subgraphs for each target node to prevent data leakage across target nodes.