Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

WATS: Calibrating Graph Neural Networks with Wavelet-Aware Temperature Scaling

Created by
  • Haebom

Author

Xiaoyang Li, Linwei Tao, Haohui Lu, Minjing Dong, Junbin Gao, Chang Xu

Outline

In this paper, we propose a novel post-calibration framework, called Wavelength-based Temperature Scaling (WATS), to address the issue that the confidence estimation of graph neural networks (GNNs) does not match the actual prediction accuracy. WATS improves the confidence estimation by assigning temperature per node using graph wavelet features. Unlike existing methods, it efficiently calibrates the confidence without information of neighboring nodes or model retraining, and our experiments on various graph structures and GNN backbones show that it achieves up to 42.3% improved Expected Calibration Error (ECE) and 17.24% reduced average calibration variance compared to existing methods.

Takeaways, Limitations

Takeaways:
We present a novel post-calibration method (WATS) that effectively solves the reliability estimation problem of GNNs.
Sophisticated node-by-node reliability compensation by leveraging the scalability and phase sensitivity of graph wavelets.
Achieve excellent performance without model retraining or accessing neighbor information.
Maintain efficient computational performance across a wide range of graph sizes and densities.
Limitations:
Further validation of the generality of the proposed method is needed, and an evaluation is needed to determine whether it is overly dependent on certain types of graph structures.
There is a need to further expand comparative analysis with other post-calibration methods (e.g., comparison using different types of calibration metrics).
Further evaluation of performance and stability when applied to actual safety-critical systems is required.
👍