Research map / Graph learning / Graph self-supervised learning
GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GOLD, a graph OOD detector that synthesizes pseudo-OOD node embeddings with a latent generative model trained adversarially against an evolving GNN and energy-based detector, without any real OOD data or pre-trained generators.
Problem
Graph OOD node detection typically needs auxiliary real OOD data (OOD exposure) or pre-trained generative models that do not exist for graphs; the paper asks how to expose a detector to OOD-like scenarios using only in-distribution data.
Approach
Alternate between training a latent generative model (LDM or VAE) to mimic the ID hidden embeddings of an evolving GNN, and training the GNN and an MLP energy-based detector to diverge energy scores between ID embeddings and the generated pseudo-OOD embeddings; the divergence implicitly pushes the generated samples away from ID data, making them pseudo-OOD.
Key results
- On Twitch, GOLD (LDM) reaches 99.46 AUROC and 1.78 FPR95, versus 33.57 FPR95 for the best real-OOD-exposed baseline GNNSafe++.
- GOLD (LDM) improves Cora AUROC to 95.84 and FPR95 to 17.83, beating non-exposed NODESafe (AUROC 94.39, FPR95 26.04).
- On Arxiv, GOLD (LDM) gets AUROC 73.90, below the real-OOD NODESafe++ (75.49), but above the non-exposed GNNSafe (71.06).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · Twitch | AUROC: 98.50 -> 99.46 | Helped |
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · Twitch | FPR95: 3.43 -> 1.78 | Helped |
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · Cora | AUROC: 95.36 -> 95.84 | Helped |
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · Cora | FPR95: 20.20 -> 17.83 | Helped |
| use GOLD with LDM instead of GNNSafe++ | Graph OOD detection · Amazon-Photo | AUROC: 98.99 -> 98.81 | Hurt |
| use GOLD with LDM instead of GNNSafe++ | Graph OOD detection · Amazon-Photo | FPR95: 2.10 -> 2.07 | Helped |
| use GOLD with LDM instead of GNNSafe++ | Graph OOD detection · Coauthor-CS | AUROC: 99.28 -> 99.00 | Hurt |
| use GOLD with LDM instead of GNNSafe++ | Graph OOD detection · Coauthor-CS | FPR95: 3.18 -> 3.16 | Helped |
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · ogbn-Arxiv | AUROC: 75.49 -> 73.90 | Hurt |
| use GOLD with LDM instead of NODESafe++ | Graph OOD detection · ogbn-Arxiv | FPR95: 75.24 -> 80.57 | Hurt |
| add adversarial training to the latent generator and detector (ablation) | GOLD · Twitch | AUROC: 84.59 -> 99.46 | Helped |
| add the MLP detector instead of using GNN energy scores (ablation) | GOLD · Twitch | AUROC: 77.70 -> 99.46 | Helped |
Surprising findings
- Using an ID-pretrained generative model without adversarial training did not improve OOD detection (w/o Adv. Twitch AUROC 84.59 vs GNNSafe++ 95.36).
- Using only the energy regulariser yields AUROC near single digits on Twitch (10.18), which the authors note challenges the assumption that OOD energies exceed ID energies.
Limitations
- Performance is limited by ID classifier accuracy on Arxiv (~50%), so pseudo-OOD generation is weaker there than real OOD exposure.
- Training cost of the LDM variant is higher than the baselines; only node-level tasks are studied.
- Only node-level OOD detection is addressed; graph-level extension is future work.
Experimental setup
Datasets: Twitch-Explicit Cora Amazon-Photo Coauthor-CS ogbn-Arxiv
Benchmarks: Twitch-Explicit Cora Amazon-Photo Coauthor-CS ogbn-Arxiv
Metrics: AUROC AUPR FPR95 ID accuracy
Baselines: MSP ODIN Mahalanobis Energy GKDE GPN GNNSafe NODESafe OE Energy FT GNNSafe++ NODESafe++
Builds on: GNNSafe NODESafe Energy-based OOD detection Latent diffusion model Variational autoencoder
Closest papers by meaning
- GOODAT: Towards Test-Time Graph Out-of-Distribution Detection (AAAI 2024 · similarity 0.90)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.90)
- Revisiting Score Propagation in Graph Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.90)
- Structural Entropy Guided Unsupervised Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.89)
- Graph Out-of-Distribution Detection Goes Neighborhood Shaping (ICML 2024 · similarity 0.89)
- Bounded and Uniform Energy-based Out-of-distribution Detection for Graphs (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.