STAD-Semi-Supervised-Text-Attributed-Graph-Distillation

The repository of Semi-Supervised Text-Attributed Graph Distillation accepted to KDD2026

STAD

Linux/macOS

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv .venv --python 3.12
source .venv/bin/activate
# Install base dependencies (CPU-only PyTorch)
uv sync
# Install CUDA-enabled PyTorch (automatically detects CUDA version)
CUDA_FULL=$(nvidia-smi | grep -i "cuda version" | grep -oE "[0-9]+\.[0-9]+" | head -1)
MAJOR=$(echo $CUDA_FULL | cut -d. -f1)
CUDA_VERSION=$([ "$MAJOR" -ge 12 ] && echo "cu121" || echo "cu118")
echo "Detected CUDA $CUDA_FULL, installing PyTorch with $CUDA_VERSION"
uv pip uninstall torch torchvision torchaudio
uv pip install torch torchvision torchaudio --index-url "https://download.pytorch.org/whl/$CUDA_VERSION"
# Verify GPU is detected
python -c "import torch; print('CUDA available:', torch.cuda.is_available(), 'CUDA version:', torch.version.cuda if torch.cuda.is_available() else 'N/A')"
# Install torch-scatter and torch-sparse
TORCH_VER=$(python -c "import torch; print(torch.__version__.split('+')[0])")
CUDA_VER=$(python -c "import torch; v='cu'+torch.version.cuda.replace('.', '') if torch.cuda.is_available() and torch.version.cuda else 'cpu'; print(v)")
uv pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-${TORCH_VER}+${CUDA_VER}.html

LLM (STAD only)

The text-synthesis step calls a Qwen API. Set the API key via environment variable QWEN_KEY, or in keys.py, or pass --qwen_api_key. Without it, running STAD will raise a clear error when synthesis is reached.

Data

Download minilmdata.zip and extract so that dataset/0_splits/{dataset_name}/ and dataset/{dataset_name}/ exist.

Method

WSD-guided distillation: dual-path encoders (GCN + MLP) with collaborative self-training, WSD-based graph sketching, and LLM synthesis for condensed node text. See models/GNN.py for SelfTrainingGNN, tagcond.py for the full pipeline.

Running

Windows (PowerShell)

# Activate virtual environment (use batch file to avoid execution policy issues)
.venv\Scripts\activate
# Run the main script
python main.py --reduction_rate 0.05 --dataset cora --tok_lim 256

Linux/macOS

source .venv/bin/activate
python main.py --reduction_rate 0.05 --dataset cora --tok_lim 256

Reference

Dataset: https://github.com/CurryTang/TSGFM LLM4TAG: https://github.com/WxxShirley/LLMNodeBed