ClusterData Parent Class

[suxuann]

The core functionality of ClusterData is to partition a large graph into multiple subgraphs.

My question is: Why does it subclass from the generic torch.utils.data.Dataset, and not torch_geometric.data.dataset.Dataset?

The rationale for subclassing from the latter class is that we can view a dataset consisting of subgraphs as a new dataset altogether, and use some of its functionalities (although admittedly, there are some differences).

I mostly followed the example in https://github.com/rusty1s/pytorch_geometric/blob/master/examples/cluster_gcn_reddit.py

Additionally, I bumped into a problem using ClusterLoader while using a for-loop over it. It seems like it only has one element even though I have specified num_parts in ClusterData.

[suxuann]

Moreover, the original paper contains methods to group clusters, and use these for training. This seems to be lacking in the current implementation. Would it be useful?

[rusty1s]

While one can view ClusterData as a dataset/list of subgraphs, it is not recommended in doing so because you will lose out on the inter-connectivity between clusters when iterating over it in mini-batches. The ClusterData is more meant as a simple interface for inspecting different subgraphs, while it is recommended to use ClusterLoader to iterate over it.

I think one could also implement it as a subclass of torch_geometric.data.Dataset, but we do not need the download and process functionality that the PyG dataset interface provides.