Research map / Graph learning / Graph self-supervised learning

Bridging OOD Detection and Generalization: A Graph-Theoretic View

NeurIPS 2024 · Graph learning › Graph self-supervised learning · approach: Graph out-of-distribution generalization · 1 citations (OpenAlex)

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 a graph-spectral framework that jointly analyzes OOD generalization (linear probing error on covariate shift) and OOD detection (ID–semantic separability) under a wild-data mixture, using a spectral contrastive surrogate loss.

Problem

Joint OOD generalization and OOD detection when unlabeled wild data mixes in-distribution, covariate-shifted OOD, and semantic-shifted OOD samples; prior work lacks a formal understanding of how wild data affects both.

Approach

Builds a graph over augmented ID-labeled and unlabeled wild samples combining self-supervised and supervised connectivity; representations come from factorizing the normalized adjacency matrix (spectral decomposition), equivalent to a spectral contrastive loss, enabling closed-form error bounds for generalization and detection.

Key results

• On SVHN semantic OOD with CIFAR-10-C covariate shift, the method reaches 86.62% OOD accuracy and 0.13% FPR95, versus 84.69% and 10.86% for Scone.
• On Textures, FPR95 drops to 12.05% vs 37.15% for Scone, a reduction of about 25.10 points.
• Averaged across five semantic-shift OOD datasets, FPR95 is reduced by 8.34% relative to Scone (as claimed in the introduction).

Measured effects reported in the paper

Surprising findings

• Methods trained with wild data act as strong OOD detectors but show notable declines in OOD generalization accuracy, indicating covariate OOD data confuses them.

Limitations

• Only covariate shift is studied for OOD generalization; other shifts such as concept shift are deferred.
• Theoretical analysis relies on a toy five-node graph with approximations (small terms set to zero).

Experimental setup

Datasets: CIFAR-10 CIFAR-10-C SVHN LSUN Places365 Textures

Benchmarks: CIFAR-10 CIFAR-10-C SVHN LSUN-C Textures Places365 LSUN-R ImageNet Office-Home

Metrics: OOD accuracy ID accuracy FPR95 AUROC

Baselines: MSP ODIN Energy Mahalanobis ViM KNN ASH ERM IRM Mixup VREx EQRM

Builds on: spectral contrastive loss (HaoChen et al.) SCONE WOODS Outlier Exposure

Closest papers by meaning

• GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.90)
• Graph Structure Extrapolation for Out-of-Distribution Generalization (ICML 2024 · similarity 0.89)
• Aggregation Hides Out-of-Distribution Generalization Failures from Spurious Correlations (NeurIPS 2025 · similarity 0.89)
• Graph Out-of-Distribution Detection Goes Neighborhood Shaping (ICML 2024 · similarity 0.89)
• The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.89)
• Double Descent Meets Out-of-Distribution Detection: Theoretical Insights and Empirical Analysis on the Role of Model Complexity (NeurIPS 2025 · similarity 0.89)