ICRL: independent causality representation learning for domain generalization
Abstract Domain generalization (DG) addresses the challenge of out-of-distribution (OOD) data; however, the reliance on statistical correlations during model development often introduces shortcut learning problems. Current approaches to mitigating these issues commonly involve the integration of cau...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-04-01
|
| Series: | Scientific Reports |
| Subjects: | |
| Online Access: | https://doi.org/10.1038/s41598-025-96357-0 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Domain generalization (DG) addresses the challenge of out-of-distribution (OOD) data; however, the reliance on statistical correlations during model development often introduces shortcut learning problems. Current approaches to mitigating these issues commonly involve the integration of causal inference, formalizing DG problems through a general structural causal model. Nevertheless, ensuring the independence of features when incorporating causal models is often overlooked, leading to spurious causal relationships. In this work, we design three independent feature modules using GAN variants (GAN, WGAN, and WGAN-GP) and select the best-performing WGAN module to integrate into the existing causal model framework, constructing an independent causal relationship learning (ICRL) model. Extensive experiments on widely used datasets demonstrate that our model, with independent causal representations, outperforms the original model in both performance and efficiency, thereby validating the effectiveness of our proposed approach. The code for ICRL can be accessed at: https://github.com/22Shao/ICRL.git . |
|---|---|
| ISSN: | 2045-2322 |