Research map / Label-efficient and robust learning / Domain adaptation and generalization
HiGDA: Hierarchical Graph of Nodes to Learn Local-to-Global Topology for Semi-Supervised 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 HiGDA, a hierarchical graph of nodes (local patch graph per image plus global graph over same-category samples) with a graph active learning strategy for semi-supervised domain adaptation, reporting state-of-the-art SSDA accuracy on Office-Home, DomainNet and VisDA2017.
Problem
Semi-supervised domain adaptation under domain shift, where models fail to capture rich feature-level and category-level representations and ignore structural relations among same-class samples.
Approach
Builds a two-level graph: a local k-NN graph over image patches (Pyramid ViG style max-relative graph convolution) for feature-level aggregation, and a global graph over image embeddings in a mini-batch connecting same-category samples for category-level aggregation; a graph active learning step adds confident target pseudo-labels to the labeled set each episode, and the design is combined with MM
Key results
- On Office-Home 3-shot, HiGDA-T+GAL+MME reaches 90.2% average accuracy versus 85.2% for HiGDA-T+MME and 79.8% for S+T baseline (ResNet-34 backbone table).
- On DomainNet, HiGDA-T+GAL+MME surpasses EFTL by up to 15.2% (1-shot) and 15.8% (3-shot) per the text; 3-shot average reported as 93.4% for HiGDA-T+GAL+MME in the table.
- On VisDA2017 3-shot, HiGDA-T+GAL+MME achieves 97.8% mean class-wise accuracy, above MCL's 87.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GAL to HiGDA-T with S+T training (ablation) | HiGDA-T · Office-Home 3-shot | accuracy: 5.3% | Helped |
| Add GAL to HiGDA-T with S+T training (ablation) | HiGDA-T · DomainNet 3-shot | accuracy: 7.1% | Helped |
| Add AAC (ablation) | HiGDA-T · Office-Home 3-shot | accuracy: 7.2% | Helped |
| Add MME (ablation) | HiGDA-T · Office-Home 3-shot | accuracy: 10.7% | Helped |
| Add AAC (ablation) | HiGDA-T · DomainNet 3-shot | accuracy: 9.5% | Helped |
| Add MME (ablation) | HiGDA-T · DomainNet 3-shot | accuracy: 11.1% | Helped |
| Replace MLP with GoG (ablation) | LoG-Tiny classifier on DomainNet · DomainNet 3-shot | accuracy: 64.5 -> 75.3 | Helped |
| Replace ResNet-34 with LoG-Tiny (ablation) | GoG classifier on DomainNet · DomainNet 3-shot | accuracy: 73.0 -> 75.3 | Helped |
| Replace PVT-Tiny with LoG-Tiny (ablation) | GoG classifier on DomainNet · DomainNet 3-shot | accuracy: 64.2 -> 75.3 | Helped |
| Replace ResNet-34 with LoG-Tiny (ablation) | MLP classifier on DomainNet · DomainNet 3-shot | accuracy: 61.6 -> 64.5 | Helped |
| Replace PVT-Tiny with LoG-Tiny (ablation) | MLP classifier on DomainNet · DomainNet 3-shot | accuracy: 57.5 -> 64.5 | Helped |
| Replace GoG with MLP (ablation) | LoG-Tiny on DomainNet · DomainNet 3-shot | accuracy: 75.3 -> 64.5 | Hurt |
Surprising findings
- Adding GAL pseudo-labels yields large gains over the S+T baseline, e.g. a 5.3% drop on Office-Home when GAL is removed under 3-shot.
- The HiGDA-T baseline without GAL already exceeds several CNN-based SSDA methods on Office-Home, for example surpassing DECOTA by 5.3% in 1-shot.
Limitations
- GoG is sensitive to noisy samples whose images do not match their annotated labels, causing cumulative errors.
- Balancing node and edge losses is critical and currently not optimized.
- Pseudo-label quality and threshold choice (tau=0.95) affect GAL performance.
Experimental setup
Datasets: Office-Home DomainNet VisDA2017
Benchmarks: Office-Home DomainNet VisDA2017
Metrics: classification accuracy (%) mean class-wise accuracy (%)
Baselines: MME CDAC DECOTA MCL SPA MME + SLA FixMME + EFTL ECACL APE ProML PVT-Tiny ResNet-34
Builds on: Pyramid ViG Minimax Entropy (MME) Adversarial Adaptive Clustering (AAC) EdgeConv-style max-relative graph convolution Graph Convolutional Network
Closest papers by meaning
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.88)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.87)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.87)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.87)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.86)
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.