KEFI: Kernel-based Feature Identification for Generalizable Classification

Abstract

To achieve satisfactory generalization performance on previously unseen domains, existing domain generalization (DG) methods often assume fixed domain-invariant features from a set of training domains for good generalization on new domains. However, this assumption can be overly strict, especially when the source domains lack shared information or when the target domains utilize information from selective source domains in a compositional manner. This leads to the natural question of how we utilize information from the source domain to the target domain in an appropriate way. In response to this challenge, we propose an innovative framework that includes an attribute-based feature extractor that captures from the source domains semantically meaningful components referred to as \textit{attributes} and a \textit{Kernel-based Attribute Identifier} that leverages kernel learning theory to define the decision boundaries for these attributes collected from the source domains. This dynamic learning approach empowers the classifier to effectively identify the learned attributes in the domains it has not encountered before. We empirically validate our method on well-established DG benchmarks, achieving competitive results compared to state-of-the-art techniques.