[Model] NASA

[wensi123]

1. Motivation and Context

This PR introduces an implementation of the NASA model from the paper: "Regularizing Graph Neural Networks via Consistency-Diversity Graph Augmentations".

The core contribution of this paper is a novel regularization method for Graph Neural Networks that improves performance and generalization in semi-supervised node classification tasks. It achieves this through:

• A new graph augmentation strategy called Neighbor Replacement (NR), which perturbs the graph structure by replacing 1-hop neighbors with 2-hop neighbors to balance consistency and diversity.
• A neighbor-constrained regularization loss (L_CR) that enforces prediction consistency among neighboring nodes in the augmented graph, effectively leveraging unlabeled data.

2. Summary of Changes

This PR adds the following components:

• Core Library Components:

  • gammagl/transforms/nr_augmentor.py: A new NR_Augmentor_GammaGL class that performs the Neighbor Replacement augmentation on a gammagl.data.Graph object.
  • gammagl/models/nasa_gcn.py: A new NASA_GCN_GammaGL model, which is a GCN backbone equipped with the custom compute_nasa_loss function that includes both the supervised cross-entropy loss and the neighbor-constrained regularization loss.
  • gammagl/transforms/__init__.py: Exported NR_Augmentor_GammaGL.
  • gammagl/models/__init__.py: Exported NASA_GCN_GammaGL.
  • gammagl/utils/nasa_utils.py: Utility functions (accuracy_tlx, compute_gcn_norm) required for the example.

• Example and Documentation:

  • examples/nasa/: A new directory for the NASA example.
  • examples/nasa/nasa_gcn_trainer.py: A complete training and evaluation script demonstrating how to use the new model and augmentor.
  • examples/nasa/README.md: A clear guide on how to run the example and understand its components.