Reproduce “uLEAD-TabPFN: Uncertainty-aware Dependency-based Anomaly Detection with TabPFN” on OpenResearch

About this paper

Sha Lu, Jixue Liu, Stefan Peters, Thuc Duy Le, Craig Xie, Lin Liu, Jiuyong Li

uLEAD-TabPFN introduces an uncertainty-aware, dependency-based anomaly detection framework for tabular data, leveraging a dependency-aligned latent space, Prior-Data Fitted Networks (TabPFN) for conditional dependency modeling, and uncertainty-aware scoring. The framework achieves top average ROC-AUC ranks on high-dimensional and medium-dimensional tabular datasets, demonstrating improved robustness and complementary anomaly detection capabilities compared to existing state-of-the-art methods.

Key insights

• Leveraging pre-trained foundation models like TabPFN through in-context inference enables robust and expressive conditional dependency modeling without the need for extensive dataset-specific training.
• Learning a latent representation explicitly aligned with dimension-wise conditional dependency estimation significantly improves the scalability and stability of anomaly detection in high-dimensional tabular data.
• Incorporating input-dependent conditional variance into anomaly scoring helps to account for heteroscedasticity in normal data, yielding more calibrated and reliable anomaly scores by reducing false positives.

Summary by alphaXiv.