NTSFormer

Source code and dataset of the paper "NTSFormer: A Self-Teaching Graph Transformer for Multimodal Isolated Cold-Start Node Classification", which is accepted by AAAI 2026.

Homepage and Paper

• Homepage (NTSFormer): https://github.com/CrawlScript/NTSFormer
• Paper Access:
  • ArXiv: https://arxiv.org/abs/2507.04870

Dataset Download

We use three public datasets: Movies from the MAGB benchmark, and Ele-fashion and Goodreads-NC from the MM-Graph benchmark. Please refer to their official repositories for download instructions:

• MAGB: https://github.com/sktsherlock/MAGB
• MM-Graph: https://github.com/mm-graph-benchmark/mm-graph-benchmark

The mmgraph directory is copied from MM-Graph repository for the dataloader.

Requirements:

• PyTorch
• torchmetrics
• DGL
• argcfg
• scikit-learn
• shortuuid
• transformers

Run

After downloading the datasets, you can run the code with the following command:

sh run_ntsformer.sh

You can simply run the script—it will raise errors indicating that certain directories do not exist. These paths show where you should place the downloaded datasets from the MAGB or MM-Graph repositories.

Cite

If you use NTSFormer in a scientific publication, we would appreciate citations to the following paper:

@misc{hu2025ntsformerselfteachinggraphtransformer,
      title={NTSFormer: A Self-Teaching Graph Transformer for Multimodal Isolated Cold-Start Node Classification}, 
      author={Jun Hu and Yufei He and Yuan Li and Bryan Hooi and Bingsheng He},
      year={2025},
      eprint={2507.04870},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2507.04870}, 
}