Mitigating Label Noise using Prompt-Based Hyperbolic Meta-Learning in Open-Set Domain Generalization

TL;DR

This work tackles Open-Set Domain Generalization under Noisy Labels (OSDG-NL) by constructing two benchmarks from PACS and DigitsDG and introducing HyProMeta, a framework that combines hyperbolic category prototypes for label-noise-aware meta-learning with a learnable new-category prompt for improved generalization to unseen classes. Prototypes are computed in hyperbolic space (Poincaré ball), with distances guiding clean/noisy sample partitioning and noisy labels relabeled to nearest prototypes; a prompt-based augmentation introduces an auxiliary out-of-distribution category to reinforce robustness. Across extensive experiments on PACS, DigitsDG, and DomainNet with symmetric and asymmetric label noise, HyProMeta achieves state-of-the-art OSCR and H-score while maintaining strong closed-set accuracy, demonstrating robust performance under diverse backbones and noise regimes. The work provides a practical, reproducible solution to the dual challenges of label noise and open-set/domain shifts, with code and benchmarks released for further research and application.

Abstract

Open-Set Domain Generalization (OSDG) is a challenging task requiring models to accurately predict familiar categories while minimizing confidence for unknown categories to effectively reject them in unseen domains. While the OSDG field has seen considerable advancements, the impact of label noise--a common issue in real-world datasets--has been largely overlooked. Label noise can mislead model optimization, thereby exacerbating the challenges of open-set recognition in novel domains. In this study, we take the first step towards addressing Open-Set Domain Generalization under Noisy Labels (OSDG-NL) by constructing dedicated benchmarks derived from widely used OSDG datasets, including PACS and DigitsDG. We evaluate baseline approaches by integrating techniques from both label denoising and OSDG methodologies, highlighting the limitations of existing strategies in handling label noise effectively. To address these limitations, we propose HyProMeta, a novel framework that integrates hyperbolic category prototypes for label noise-aware meta-learning alongside a learnable new-category agnostic prompt designed to enhance generalization to unseen classes. Our extensive experiments demonstrate the superior performance of HyProMeta compared to state-of-the-art methods across the newly established benchmarks. The source code of this work is released at https://github.com/KPeng9510/HyProMeta.

Figures (4)

• Figure 1: Illustration of the Open-Set Domain Generalization under Noisy Label (OSDG-NL) task. Different shapes (e.g., circles, triangles, squares) symbolize distinct domains, while their respective colors represent various categories. A subset of these samples, outlined in red, indicates the presence of label noise. The objective of this task is to train a model capable of leveraging robust priors derived from source domains, despite label noise, to achieve precise classification of known categories while assigning low confidence scores to novel, unseen categories in a new target domain.
• Figure 2: Overview of the proposed HyProMeta framework for Open-Set Domain Generalization under Noisy Labels (OSDG-NL). The model first computes hyperbolic category-wise prototypes from source domain data and uses the mode of hyperbolic distances to partition samples into clean and noisy subsets. Clean samples are used directly for meta-training, while noisy samples are relabeled based on proximity to prototypes. A learnable prompt is introduced for category-aware augmentation, where mixed samples from different classes are combined with prompt patches and assigned to an auxiliary out-of-distribution class. The framework iteratively updates through meta-train and meta-test phases, enabling robust generalization to unseen categories and domains despite label noise.
• Figure 3: T-SNE visualization van2008visualizing of learned representations on PACS using ResNet18 he2016deep under symmetric label noise with a ratio of $20\%$ when we use Photo and Art Painting as test domains, respectively. D indicates the target domain and LNSR indicates the label noise setting.
• Figure 4: Confidence score visualization of learned representations on PACS using ResNet18 he2016deep under symmetric label noise with a ratio of $50\%$ and asymmetric label noise with a ratio of $50\%$. D indicates the target domain and LNSR indicates the label noise setting.