Research map / Label-efficient and robust learning / Domain adaptation and generalization
Probability-Polarized Optimal Transport for Unsupervised 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 Probability-Polarized Optimal Transport (PPOT), which adds a margin-based polarization regularizer (with dynamic, label-informed slack thresholds, DPP) to entropic OT so the transport plan has explicit intra/inter-class structure, and applies it to unsupervised domain adaptation.
Problem
Standard entropy-regularized OT yields dense transport plans that blur intra- and inter-class correspondences and add uncertainty, hurting class-structured distribution alignment in unsupervised domain adaptation.
Approach
Adds a probability polarization regularizer that pushes coupling values away from an intermediate threshold (toward large intra-class or small inter-class mass), with dynamic margins adjusted by source labels and target pseudo-labels, solved via a generalized conditional gradient algorithm.
Key results
- PPOT achieves 93.5% average accuracy on Office-31 and 91.9% on ImageCLEF, the best among compared methods.
- On DomainNet, PPOT reaches 34.0% average accuracy, at least 1.6% above other methods.
- On ImageCLEF P→I, PPOT's OT plan has 91.59% intra-class mass versus 75.11% for entropic OT (OT-IT).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PPOT instead of RWOT | UDA on Office-31 · Office-31 average | accuracy: 90.8 -> 93.5 | Helped |
| use PPOT instead of DDW-OT | UDA on ImageCLEF · ImageCLEF average | accuracy: 90.7 -> 91.9 | Helped |
| use PPOT instead of DCAN | UDA on DomainNet · DomainNet average | accuracy: 32.4 -> 34.0 | Helped |
| replace the PP regularizer with the DPP regularizer (ablation) | PPOT on ImageCLEF · ImageCLEF average | accuracy: 87.8 -> 88.8 | Helped |
| add Lt to the DPP model using Lot only (ablation) | PPOT on ImageCLEF · ImageCLEF average | accuracy: 86.9 -> 88.8 | Helped |
| add Lot to the DPP model using Lt only (ablation) | PPOT on ImageCLEF · ImageCLEF average | accuracy: 88.1 -> 88.8 | Helped |
| pre-train on the source domain before using pseudo-labels (ablation) | PPOT · ImageCLEF P→I | training stability | Helped |
Limitations
- Polarization thresholds rely on source labels and target pseudo-labels, so errors in pseudo-labels can cause wrong polarization directions.
- Evaluated only on single-source UDA image classification benchmarks.
Experimental setup
Datasets: Office-31 ImageCLEF DomainNet
Benchmarks: Office-31 ImageCLEF DomainNet
Metrics: classification accuracy intra-class mass LDA discriminant ratio training time and iterations
Baselines: DANN CDAN+E BCDM DALN DeepJDOT RWOT DDW-OT PCT DMP PGFL SCDA DCAN
Builds on: entropy-regularized optimal transport (Cuturi 2013) polarization regularization (Chen et al. 2021) generalized conditional gradient DeepJDOT RWOT
Closest papers by meaning
- Prototypical Partial Optimal Transport for Universal Domain Adaptation (AAAI 2023 · similarity 0.89)
- Reliable Weighted Optimal Transport for Unsupervised Domain Adaptation (CVPR 2020 · similarity 0.88)
- Unbalanced minibatch Optimal Transport; applications to Domain Adaptation (ICML 2021 · similarity 0.88)
- Incorporating Importance Weighting in Optimal Transport Based Domain Alignment (ICML 2026 · similarity 0.88)
- Enhanced Transport Distance for Unsupervised Domain Adaptation (CVPR 2020 · similarity 0.88)
- Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport (ICML 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.