Research map / Label-efficient and robust learning / Domain adaptation and generalization
Category-Aware Active Domain Adaptation
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 category-aware active domain adaptation that uses influence functions to select target samples that most improve the recognition of a specified category, without hurting other categories.
Problem
Existing active domain adaptation methods improve overall accuracy but concentrate queries on easy-to-improve categories, leaving challenging categories unimproved or degraded.
Approach
Binarize labels into a target category versus the rest, train a convex logistic-regression surrogate, compute influence of labeled source samples on a validation loss over queried target samples, train a regressor to predict influence for unlabeled target samples, and query high-influence ones per category.
Key results
- On Office-Home the method averages 81.01% (Ar→Cl through Cl→Rw) and 78.64% on the other six tasks, the best on 11 of 12 tasks, beating ISAL and Always One.
- On DomainNet-126 the method averages 57.78% over six tasks vs 53.33% for Always One (best baseline) per Table 4; on VisDA-2017 S→R it reaches 66.68% vs 63.25%.
- Ablation on Rw→Ar: influence-only 72.74%, full model with re-weighting and diversity sampling 79.15%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use category-aware influence-based active learning instead of Always One | DANN-based category-aware active domain adaptation · Office-Home | average per-category accuracy: 79.32% across 12 tasks; Always One obtains an average of 75.81% in 12 tasks | Helped |
| use category-aware influence-based active learning instead of Always One | DANN-based category-aware active domain adaptation · DomainNet-126 | average per-category accuracy: 57.78 vs 53.33 | Helped |
| use category-aware influence-based active learning instead of Always One | DANN-based category-aware active domain adaptation · VisDA-2017, S→R | per-category accuracy: 66.68 vs 63.25 | Helped |
| add influence-based selection to DANN (ablation) | category-aware active domain adaptation · Office-Home, Rw→Ar | accuracy: 52.28 ± 3.63 -> 72.74 ± 5.41; 20.46% | Helped |
| add data re-weighting to influence-based selection (ablation) | category-aware active domain adaptation · Office-Home, Rw→Ar | accuracy: 72.74 ± 5.41 -> 76.52 ± 6.37; 2.74% | Helped |
| add diversity sampling to influence-based selection (ablation) | category-aware active domain adaptation · Office-Home, Rw→Ar | accuracy: 72.74 ± 5.41 -> 74.36 ± 3.18; 1.64% | Helped |
| combine data re-weighting and diversity sampling with influence-based selection (ablation) | category-aware active domain adaptation · Office-Home, Rw→Ar | accuracy: 72.74 ± 5.41 -> 79.15 ± 2.92; 6.41% | Helped |
| exclude queried samples with negative estimated influence on Non-One data (ablation) | category-aware active domain adaptation · Office-Home, Cl→Pr | Non-One category accuracy | Helped |
| train using queried One samples (ablation) | category-aware active domain adaptation · Office-Home, Pr→Ar | target-category accuracy | Helped |
| train using queried Non-One samples (ablation) | category-aware active domain adaptation · Office-Home, Pr→Ar | target-category accuracy | Helped |
Surprising findings
- Always One, which simply queries samples most likely in the target category, is the strongest baseline, outperforming general active DA methods.
- Re-training with only Non-One queried samples still benefits per-category accuracy.
Limitations
- Binarization of labels per category can cause class imbalance in the surrogate training data.
- Improvements are smaller on large-scale datasets due to limited query budget.
- Influence estimates rely on a convex logistic regression surrogate rather than the actual non-convex DA classifier.
Experimental setup
Datasets: Office-Home DomainNet-126 VisDA-2017
Benchmarks: Office-Home DomainNet-126 VisDA-2017
Metrics: per-category accuracy accuracy change on non-target categories
Baselines: Random CLUE DUC ISAL Always One DANN
Builds on: influence function (Koh & Liang 2017) DANN CLUE ISAL
Closest papers by meaning
- ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation (CVPR 2025 · similarity 0.91)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.90)
- Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments (AAAI 2024 · similarity 0.88)
- Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.88)
- COSDA: Counterfactual-based Susceptibility Risk Framework for Open-Set Domain Adaptation (ICML 2025 · similarity 0.88)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · 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.