Research map / Label-efficient and robust learning / Domain adaptation and generalization
Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm
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 open-set heterogeneous domain adaptation (OSHeDA) with theoretical learning bounds and RL-OSHeDA, a representation-learning method using MMD alignment, a non-negative open-set risk estimator, and 2-stage pseudo-labeling, evaluated on 7 datasets (56 DA tasks).
Problem
Domain adaptation when source and target have heterogeneous feature spaces and the target label space contains unknown classes absent from the source (open-set), where existing HeDA methods fail.
Approach
Derives JS-divergence-based error bounds for open-set heterogeneous DA and turns them into a training objective: align known-class representations (marginal and label-conditioned, via MMD), separate unknown from known representations, and add a non-negative open-set risk term, with pseudo-labels from a decoupled 2-stage schedule.
Key results
- RL-OSHeDA averages HOS 67.21 across the 7 datasets, versus 59.89 for the best baseline SCT (Table 1).
- On Multilingual Reuters, the full method reaches HOS 65.39 versus 58.33 with no proposed components (ablation).
- Friedman test with Nemenyi post-hoc over 56 DA tasks reports RL-OSHeDA significantly better than baselines (P < 0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RL-OSHeDA instead of the strongest baseline, STN | OSHeDA · CIFAR10 & ILSVRC2012 | HOS: 61.59$\pm$0.72 -> 72.33$\pm$0.70 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, SCT | OSHeDA · ImageCLEF-DA | HOS: 58.76$\pm$2.05 -> 63.98$\pm$2.04 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, SCT | OSHeDA · Multilingual Reuters Collection | HOS: 61.17$\pm$0.94 -> 65.39$\pm$0.91 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, OPDA | OSHeDA · NUSWIDE & ImageNet | HOS: 71.06$\pm$1.44 -> 80.01$\pm$1.30 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, SCT | OSHeDA · Office & Caltech256 | HOS: 75.72$\pm$2.14 -> 78.18$\pm$2.05 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, SCT | OSHeDA · Wikipedia | HOS: 58.41$\pm$2.01 -> 63.10$\pm$1.89 | Helped |
| Use RL-OSHeDA instead of the strongest baseline, OPDA | OSHeDA · PTB-XL | HOS: 31.47$\pm$1.22 -> 47.48$\pm$1.25 | Helped |
| Add 2-stage learning (ablation) | RL-OSHeDA · Multilingual Reuters Collection | HOS: 59.40 -> 65.39 | Helped |
| Add Segregate ($L_{seg}$) (ablation) | RL-OSHeDA · Multilingual Reuters Collection | HOS: 61.92 -> 65.39 | Helped |
| Add Align ($L_{inv}$) (ablation) | RL-OSHeDA · Multilingual Reuters Collection | HOS: 58.23 -> 65.39 | Helped |
| Add OSD ($L_{osd}$) (ablation) | RL-OSHeDA · Multilingual Reuters Collection | HOS: 59.96 -> 65.39 | Helped |
| Add Align, Segregate, OSD, and 2-stage learning (ablation) | RL-OSHeDA · Multilingual Reuters Collection | HOS: 58.33 -> 65.39 | Helped |
Surprising findings
- Several HeDA baselines (SCT, SSAN, STN) that were not designed for open-set perform better than OPDA and KPG, which are open-set-aware methods.
- KPG could not be run on PTB-XL (reported N/A), and its SVM-on-transported-data approach underperforms.
Limitations
- Theoretical bounds assume bounded loss and rely on JS divergence estimated via adversarial learning, with an assumed sufficient representation condition.
- Requires a small labeled target set; performance depends on pseudo-label quality, and the 2-stage schedule adds a threshold hyperparameter T.
- Class proportion lambda must be estimated with positive-unlabeled techniques.
Experimental setup
Datasets: CIFAR10 ILSVRC2012 Wikipedia Multilingual Reuters Collection NUSWIDE ImageNet Office Caltech256 ImageCLEF-DA PTB-XL
Benchmarks: CIFAR10 & ILSVRC2012 ImageCLEF-DA Multilingual Reuters Collection NUSWIDE & ImageNet Office & Caltech256 Wikipedia PTB-XL
Metrics: HOS OS* UNK accuracy
Baselines: DS3L KPG OPDA PL SCT SSAN STN SL
Builds on: Heterogeneous domain adaptation (SSAN/STN/SCT) open-set domain adaptation (OPDA) DS3L KPG positive-unlabeled learning non-negative risk estimator
Closest papers by meaning
- COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.90)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.88)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.88)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.88)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.88)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.88)
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.