Research map / Label-efficient and robust learning / Domain adaptation and generalization
Target-Free Domain Adaptation through Cross-Adaptation (Student Abstract)
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 "cross-adaptation," a target-free domain adaptation method that transforms each training dataset toward the others using off-the-shelf domain adaptation methods, then trains a general model for SARS-CoV-2 detection from complete blood count data.
Problem
Merging CBC datasets from different hospitals for SARS-CoV-2 detection can harm performance because population characteristics differ, and unseen target-domain labels are unavailable for adaptation.
Approach
Iteratively treat each training dataset as source and the rest as targets, transform the source with a standard domain adaptation method (KMM, KLIEP, TrAdaBoost), concatenate all transformed datasets, and train a general classifier that is claimed to generalize to unseen domains without target labels.
Key results
- Cross-adaptation raised F1 for KNN from 55% to 73%, Decision Tree from 65% to 73%, and XGBoost from 62% to 73%.
- Across five classifiers, cross-adaptation improved average F1 by roughly 10 percentage points per the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train KNN on source datasets transformed with cross-adaptation instead of untransformed source datasets | KNN · CBC SARS-CoV-2 detection (Ethiopia and Spain) | F1 score: 55% -> 73% | Helped |
| Train Decision Tree on source datasets transformed with cross-adaptation instead of untransformed source datasets | Decision Tree · CBC SARS-CoV-2 detection (Ethiopia and Spain) | F1 score: 65% -> 73% | Helped |
| Train Random Forest on source datasets transformed with cross-adaptation instead of untransformed source datasets | Random Forest · CBC SARS-CoV-2 detection (Ethiopia and Spain) | F1 score: 62% -> 70% | Helped |
| Train XGBoost on source datasets transformed with cross-adaptation instead of untransformed source datasets | XGBoost · CBC SARS-CoV-2 detection (Ethiopia and Spain) | F1 score: 62% -> 73% | Helped |
| Train MLP on source datasets transformed with cross-adaptation instead of untransformed source datasets | MLP · CBC SARS-CoV-2 detection (Ethiopia and Spain) | F1 score: 65% -> 72% | Helped |
Limitations
- Evaluated only on a single task (SARS-CoV-2 CBC detection) and tabular features, with plans for other modalities left as future work.
- Results are reported as averages over three adaptation methods without variance or statistical tests shown.
- Only 2 held-out test hospitals; closed-source dataset limits reproducibility.
Experimental setup
Datasets: SARS-CoV-2 CBC datasets from Brazil, Italy, Poland, Ethiopia, Spain (8 public, 1 closed-source)
Benchmarks: SARS-CoV-2 CBC datasets from Ethiopia and Spain (held-out hospitals)
Metrics: F1 score
Baselines: KNN Decision Tree Random Forest XGBoost MLP
Builds on: Kernel Mean Matching KLIEP TrAdaBoost
Closest papers by meaning
- Learning Cross-Domain Landmarks for Heterogeneous Domain Adaptation (CVPR 2016 · similarity 0.84)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.83)
- LEAD: Learning Decomposition for Source-free Universal Domain Adaptation (CVPR 2024 · similarity 0.82)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.82)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.82)
- Leveraging Normalization Layer in Adapters with Progressive Learning and Adaptive Distillation for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.82)
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.