<ul><li>Multi-source domain adaptation (MSDA) aims to learn a label prediction function for an unlabeled target domain by leveraging labeled data from multiple source domains and unlabeled data from the target domain.</li><li>Conventional MSDA approaches rely on covariate shift or conditional shift paradigms, assuming a consistent label distribution across domains. However, this limits their applicability in real-world scenarios where label distributions vary across domains.</li><li>To address this limitation, a new paradigm called latent covariate shift (LCS) is proposed, introducing greater variability and adaptability across domains. It allows for recovering the latent cause of the label variable, referred to as the latent content variable.</li><li>The proposed MSDA method based on LCS achieves exceptional performance on both simulated and real-world datasets by learning the label distribution conditioned on the identifiable latent content variable, accommodating substantial distribution shifts.</li></ul>

Latent Covariate Shift: Unlocking Partial Identifiability for Multi-Source Domain Adaptation

Discover more