Reproduce “COD: Learning Conditional Invariant Representation for Domain Adaptation Regression” on OpenResearch

About this paper

Hao-Ran Yang, Chuan-Xian Ren, You-Wei Luo

Researchers at Sun Yat-Sen University introduced Conditional Operator Discrepancy (COD), a kernel-based metric, to enable conditional invariant representation learning for Domain Adaptation Regression (DAR) with continuous labels. This method consistently achieved lower Mean Absolute Error on DAR benchmarks, outperforming existing state-of-the-art techniques.

Key insights

• For Domain Adaptation Regression, achieving conditional invariant representations is theoretically sufficient to bound the cross-domain generalization error, providing a strong basis for conditional alignment.
• The 'continuity problem' in DAR can be effectively addressed by characterizing and matching conditional distributions through their statistical moments in RKHS, as demonstrated by the Conditional Operator Discrepancy (COD) metric.
• Preserving discriminability within domains is crucial during conditional alignment; the modified COD term shows that explicitly maximizing intra-class similarity within each domain improves overall representation quality and performance.

Summary by alphaXiv.