Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization

2025-05-29

AI Summary

Deep learning models often suffer performance degradation under test-time distribution shifts, and while multi-source domain generalization (MDG) mitigates this, it requires costly, labor-intensive multi-domain datasets—making it inapplicable to single-source domain (SDG) settings. Method: We propose pseudo-multi-source domain generalization (PMDG), a framework that synthesizes multiple semantically consistent yet stylistically diverse pseudo-domains from a single source domain via style transfer (e.g., AdaIN) and diversity-aware data augmentation, enabling direct application of existing MDG algorithms (e.g., ERM, CORAL, Mixup-MDG). Contribution/Results: Evaluated on our benchmark PseudoDomainBed, PMDG significantly improves generalization from a single source; with sufficient pseudo-domains, it matches or surpasses real multi-source MDG performance; and MDG and PMDG exhibit strong positive correlation in efficacy. Code and benchmark are open-sourced to standardize SDG evaluation.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Diffusion Models for Vision
• Application Domains: Misinformation & Fake News

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain dataset that can be used with existing MDG algorithms. Through extensive experiments with PseudoDomainBed, our modified version of the DomainBed benchmark, we analyze the effectiveness of PMDG across multiple datasets and architectures. Our analysis reveals several key findings, including a positive correlation between MDG and PMDG performance and the potential of pseudo-domains to match or exceed actual multi-domain performance with sufficient data. These comprehensive empirical results provide valuable insights for future research in domain generalization. Our code is available at https://github.com/s-enmt/PseudoDomainBed.

Problem

• Enhancing model performance across diverse data distributions
• Overcoming single-source limitations via pseudo multi-domain generation
• Enabling MDG algorithms in practical single-source settings

Innovation

• Generates pseudo-domains via style transfer
• Uses data augmentation for synthetic datasets
• Enables MDG algorithms in SDG settings