Add Support for HeteroGraphSAINTSampler in Heterogeneous GraphSAINT

[Gauravp2104]

🚀 The feature, motivation and pitch

Hi PyTorch Geometric Team,

We're using the torch_geometric.loader module extensively in our project for a Twitter bot detection task using Heterogeneous GNNs, and we would like to propose changes by adding a HeteroGraphSAINTSampler class in the loader, along with an extended addition of a HeteroGraphSAINTNodeSampler subclass.

Features

• Uses edge connectivity to determine valid node samples for each node type
• Reducing redundancy in the sampled nodes using torch.unique()

Key Considerations

• The sampler first collects source nodes based on adjacency matrices before applying any selection criteria.
• If no nodes are found for a specific type, it prevents unnecessary computations and avoids errors.

Potential API

from torch_geometric.loader import HeteroGraphSAINTNodeSampler
train_loader = HeteroGraphSAINTNodeSampler(
    data,
    batch_size={"user": 100, "tweet": 200, "list": 100},
    num_steps=1,
    sample_coverage=10,
    num_workers=0,
)

Motivation

• Improve efficiency by only selecting nodes relevant to the task, avoiding unnecessary computations.
• Ensures that training samples are well-distributed and diverse across the graph structure.

We are interested in discussing this further and contributing to testing to evaluate the feasibility of this feature. Looking forward to hearing from your team soon!

Alternatives

No response

Additional context

No response