Causal domain generalization

In the digital age, where machine learning models are ubiquitous, these models tend to rely on several assumptions to achieve high accuracy. The i.i.d. assumption, independent and identically distributed assumption, states that the training and test data are sampled from the same distribution. On the other hand, real-world scenarios are more dynamic, with training and test data not always coming from the same distribution. In such cases, models based on the i.i.d assumption perform poorly in generalization. In recent years, concerted efforts have addressed this shortcoming and led to the development of domain generalization methods . These methods seek to identify and rely on a stable subset of features resistant to distribution shifts. Many generalization approaches use causal theories to capture this invariance because causality and invariance are inextricably linked. Causality is leveraged differently depending on which part of the model pipeline we are at. In this chapter, we will learn how causality-aware domain generalization methods differ from traditional domain generalization methods, how and when causality is used to infer invariant features, and how these methods have been applied to vision, graphs, and text.