Pseudo-Labeling Domain Adaptation Using Multi-Model Learning
With the constant growth of state-of-the-art models, obtaining sufficient labeled data to train these models for specific domains has become increasingly costly. Domain adaptation methods offer a potential solution to enhance model performance in new, unseen domains while minimizing the need for man...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10909469/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the constant growth of state-of-the-art models, obtaining sufficient labeled data to train these models for specific domains has become increasingly costly. Domain adaptation methods offer a potential solution to enhance model performance in new, unseen domains while minimizing the need for manual annotation of target domain. Despite recent advances in using pseudo-labeling for domain adaptation, significant challenges remain in maximizing the effectiveness of pseudo-labeling, particularly when aiming to create informative and interpretable representations from pseudo-labels. To address these challenges, we introduce the method Pseudo-labeling Domain Adaptation (PDA), which leverages pseudo-labels generated by multiple models to create a robust cross-domain representation. Additionally, to further mitigate the domain-shift problem, we propose a novel method called UMAP Domain Adaptation (UMAP DA), a UMAP-based technique that allows for connections only between nodes from different domains. We use these representations to construct a heterogeneous bipartite graph, where a neural network is employed for final classification. Experiments on six different datasets show an average F1-score improvement of 8 points, measuring the harmonic mean of precision and recall, compared to existing methods in the literature. The proposed method enhances both performance and interpretability, offering a new direction for cross-domain learning with pseudo-labels. |
|---|---|
| ISSN: | 2169-3536 |