TFM4GAD: Tabular Foundation Models are Strong Graph Anomaly Detectors

This repository contains the official implementation of the paper:

Tabular Foundation Models are Strong Graph Anomaly Detectors
Yunhui Liu, Tieke He, Yongchao Liu, Can Yi, Hong Jin, Chuntao Hong
Nanjing University & Ant Group, China

TFM4GAD is a simple yet effective framework that adapts Tabular Foundation Models (TFMs), such as LimiX and TabPFN, for Graph Anomaly Detection (GAD). It bridges the structural gap between graphs and tabular data via graph-to-table flattening, enabling in-context anomaly detection without any graph-specific training.

---

📂 Project Structure

TabPFN
├─ckpts                           # Directory for pretrained model checkpoints (.ckpt)
└─tfm4gad                         # Core implementation of the TFM4GAD framework
        augment.py                # Graph-to-tabular feature augmentation (Laplacian, structural, neighborhood)
        dataloader.py             # Dataset loading and preprocessing for GADBench datasets
        icl.conf.yaml             # Configuration file for datasets, parameters, and in-context learning setup
        main.py                   # Main script to run experiments and inference
        misc.py                   # Utility functions (metrics, logging, random seeds, etc.)

Each backbone (LimiX and TabPFN) provides a self-contained implementation of the TFM4GAD framework with identical structure.

---

🚀 Getting Started

1. Prepare Datasets

We use four public datasets from GADBench: Amazon, YelpChi, T-Finance, and T-Social. Download them from Google Drive and unzip into TFM4GAD/datasets.

Dataset	Domain	#Nodes	#Edges	#Dim.	Anomaly
Amazon	Co-review	11,944	4,398,392	25	9.5%
YelpChi	Co-review	45,954	3,846,979	32	14.5%
T-Finance	Transaction	39,357	21,222,543	10	4.6%
T-Social	Social	5,781,065	73,105,508	10	3.0%

---

2. Prepare Tabular Foundation Models

🧮 LimiX

1. Download the LimiX source code from limix-ldm/LimiX (we use commit a50777d);
2. Merge it with the directory TFM4GAD/LimiX;
3. Download pretrained checkpoints from Hugging Face:

• LimiX-16M.ckpt
• LimiX-2M.ckpt
• and place them under TFM4GAD/LimiX/ckpts.

License Notice: Code under Apache-2.0 License. Weights are available for academic research, with commercial use requiring authorization.

---

🧮 TabPFN

1. Install the library pip install tabpfn;

2. Download pretrained checkpoints from Hugging Face:

   • tabpfn-v2-classifier.ckpt
   • tabpfn-v2.5-classifier-v2.5_default.ckpt
   • and place them underTFM4GAD/TabPFN/ckpts.

License Notice:

• TabPFN-v2: Prior Labs License (Apache 2.0 + attribution)
• TabPFN-v2.5: Non-commercial research license. For enterprise use, contact sales@priorlabs.ai.

---

3. Run Experiments

All configurations are stored in icl.conf.yaml.

# Using LimiX backbone
cd TFM4GAD/LimiX/tfm4gad
python main.py --dataset amazon --model_ckpt "../ckpts/LimiX-16M.ckpt"

# Using TabPFN backbone
cd TFM4GAD/TabPFN/tfm4gad
python main.py --dataset amazon --model_ckpt "../ckpts/tabpfn-v2-classifier.ckpt"

---

✨ Highlights

• One-for-All GAD: No dataset-specific training — generalizes across graph domains.
• Graph-to-Table Flattening: Enriches node features with Laplacian embeddings, structural metrics (Degree, PageRank), and anomaly-sensitive neighborhood features.
• In-Context Learning: Leverages pretrained TFMs to infer anomaly patterns directly from a few labeled samples.
• Strong Empirical Results: Outperforms state-of-the-art graph anomaly detectors on Amazon, YelpChi, T-Finance, and T-Social, achieving up to +10% AUPRC gain over trained baselines.
• No Fine-Tuning Needed: Zero training or gradient updates — pure inference-based anomaly detection.

---

🙏 Acknowledgements

We thank the authors of GADBench for providing datasets and baselines that form the foundation of this work.

We also thank the developers of LimiX and TabPFN for their remarkable contributions to tabular foundation modeling.

---

📜 License

• Code: Apache-2.0
• LimiX Weights: Subject to LimiX Model License
• TabPFN Weights: Subject to Prior Labs License