Domain Generalization via Aligning Class Hierarchical Structure in Quotient Space
Jiao ZHAO , Junbiao CUI , Qin YUE , Bin ZHENG , Jiye LIANG
Domain generalization aims to learn a robust model from multiple source domains that can perform well on target domain. The key to domain generalization is to learn representations that are both domain-invariant and class-discriminative. In most existing methods, these two objectives are modeled in a decoupled manner. This leads to a trade-off where optimizing for domain invariance inevitably impairs class discriminability, and vice versa. To address this issue, we propose an intrinsically consistent optimization objective. Our method achieves domain-invariant representations by aligning class hierarchical structure in the quotient space, thereby enabling these two objectives to be mutually reinforcing. Specifically, our approach is two-stage. First, we employ a Fuzzy Learning Machine to estimate the domain-agnostic fuzzy similarity relations among classes to construct hierarchical structure. Second, this hierarchical structure guides the alignment of class hierarchical structure of each source domain in a quotient space. Comprehensive experiments across multiple benchmarks demonstrate that our method achieves superior generalization performance compared to state-of-the-art methods.
Domain generalization / Fuzzy learning machine / Class hierarchical structure / Quotient space
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(8137KB)
