Instrumental Variable-Driven Domain Generalization with Unobserved Confounders

Domain generalization (DG) aims to learn from multiple source domains a model that can generalize well on unseen target domains. Existing DG methods mainly the representations with invariant marginal distribution of input features, however, invariance conditional labels given features is more essential for unknown domain prediction. Meanwhile, existing unobserved confounders which affect and simultaneously cause spurious correlation hinder learning relationship contained in distribution. Interestingly, causal view data generating process, we find one are valid instrumental variables other Inspired by this finding, propose an variable-driven method (IV-DG) removing bias two-stage learning. In first stage, it learns another domain. second estimates predicting learned Theoretical analyses simulation experiments show accurately captures relationship. Extensive real-world datasets demonstrate IV-DG yields state-of-the-art results.