Identifying Latent Causal Content for Robust Domain Adaptation under Significant Label Distribution Shifts

The core message of this work is that by introducing a latent causal model with a latent content variable, it is possible to identify this latent content variable up to block identifiability, which enables learning an invariant conditional distribution of labels given the latent content. This provides a principled way to achieve robust domain adaptation, especially in the presence of significant label distribution shifts across domains.