Research map / Label-efficient and robust learning / Domain adaptation and generalization

COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation

ICML 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Source-free domain adaptation · 0 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 COSDA, a counterfactual susceptibility-risk framework for open-set domain adaptation with a risk upper bound, mutual-information-based contrastive feature alignment, and virtual multi-unknown prototype pseudo-labeling.

Problem

Open-set domain adaptation must handle both domain shift and unknown target categories, yet there is little causal, theoretically grounded risk framework for OSDA.

Approach

Models a susceptibility risk (probability of counterfactual via Pearl's susceptibility) estimated by an intervention module, derives a computable upper bound for OSDA risk, enforces the Exogeneity assumption via a mutual-information-based contrastive alignment (CFA), and labels target samples with multiple virtual unknown prototypes (VMP) instead of one unknown class.

Key results

• On Image-CLEF, COSDA achieves 84.3% average HOS, a 2.9% gain over the prior state-of-the-art ANNA.
• On Office-31, COSDA reaches 92.6% average HOS across six tasks.
• With a CLIP backbone, COSDA-CLIP reaches 73.9% HOS on DomainNet (outperforming CROW by 14.9%) and 78.4% HOS on VisDA (outperforming CROW by 9.2%).

Measured effects reported in the paper

Surprising findings

• Removing the target-side risk term Lt drives UNK accuracy to 0, so the model cannot recognize unknown classes at all.
• The simple logit-based MLP pseudo-labeling yields 0.0 UNK accuracy in ablation, showing it has essentially no unknown-class discrimination.

Limitations

• Relies on approximations (variational inference, Hoeffding/Markov/Jensen bounds) so the theoretical bound is loose in practice.
• Requires a number of virtual unknown prototypes U set equal to K, a heuristic choice.
• Results on DomainNet and VisDA are only reported with CLIP backbone, and ablations are limited to Office-Home.

Experimental setup

Datasets: Office-Home Office-31 Image-CLEF DomainNet VisDA synthetic causal dataset

Benchmarks: Image-CLEF Office-31 Office-Home DomainNet VisDA

Metrics: OS* UNK HOS OS

Baselines: OSNN OSBP STA ROS DAOD ANNA DADO UAN GLC GLC++ OSLPP UNIOT

Builds on: Pearl probability of susceptibility OSBP ANNA VIB stability loss from Yang et al. 2024

Closest papers by meaning

• Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.90)
• Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.88)
• Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.88)
• CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.88)
• Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.87)
• Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.87)