Reproduce “GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Danny Wang, Ruihong Qiu, Guangdong Bai, Zi Huang
The GOLD framework from The University of Queensland introduces an implicit adversarial learning approach for graph out-of-distribution (OOD) detection. This method generates pseudo-OOD samples solely from in-distribution data, achieving competitive OOD detection performance against real OOD exposure methods without requiring any auxiliary OOD data or pre-trained graph generative models.
Key insights
- Pseudo-OOD samples can be effectively synthesized directly from in-distribution (ID) data through an implicit adversarial learning process, eliminating the need for auxiliary OOD data or pre-trained graph generative models.
- The dynamic interaction between the evolving GNN encoder/detector and the latent generative model ensures that synthesized pseudo-OOD samples remain relevant and effective for training, even as the in-distribution representation shifts.
- A robust OOD detector, utilizing uncertainty loss and a novel divergence regularisation on transformed energy scores, effectively maximizes the separation between ID and pseudo-OOD data distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.