GraphIFE

GraphIFE is research code for long-tailed node classification. It includes the GraphIFE training pipeline and GCN/GATv2/GraphSAGE backbones for Planetoid, Amazon, and Coauthor-CS datasets.

• main.py: training entry point.
• args.py and config.yaml: command-line and default configuration.
• data_utils.py: dataset loading, long-tail split construction, optimizers, and result reporting.
• models/, nets/, and losses/: GraphIFE components, GNN backbones, and loss functions.

The current source tree excludes datasets, TensorBoard logs, result files, parameter sweeps, and historical experiment archives.

Use Python 3.9 or later. Install a PyTorch build compatible with your CUDA runtime (or a CPU-only build) first, then install the remaining dependencies:

pip install -r requirements.txt

torch-scatter and torch-sparse contain compiled extensions. If the standard installation does not provide a wheel for your PyTorch/CUDA combination, install the matching wheels before rerunning the command above.

The 27 settings listed in config.yaml are the complete public parameter set. Command-line options override values from the selected YAML file. Unknown YAML keys and unsupported command-line options are rejected. Boolean options accept true and false. Datasets are downloaded to ./data on first use.

Training uses Beta(2, 2) feature mixing and selects the best epoch by validation accuracy. Neighbor duplication is used during warmup, followed by neighbor sampling. The number of synthetic nodes follows the class-balancing rule, and task losses use dynamic weighting based on their ratios between epochs.

python main.py --config config.yaml

For a lightweight CPU smoke run, override the relevant options:

python main.py --dataset Cora --net GCN --device cpu --epochs 2 --repetitions 1

To write a summary to result/ or TensorBoard logs to runs/, opt in with --write true and/or --verbose true.