Pattern classification in unseen environments with incomplete multi-source domain generalization

AbstractPattern classification in unseen environments is challenging due to distribution shifts caused by variations in acquisition conditions or sensor characteristics. Domain generalization (DG) addresses this issue by extracting robust representations that remain consistent across multiple source domains, enabling models to generalize to previously unseen target domains. In practice, however, limitations in data collection often result in incomplete class coverage within source domains, producing biased distributions that impede effective representation learning. To address this, we propose an Incomplete Multi-source Domain Generalization (IMDG) method. Specifically, a Cross-domain Class Completion Module (CCCM) is designed to synthesize missing classes by mutually leveraging auxiliary knowledge across heterogeneous source domains, effectively bridging category-level feature gaps. A Distance-Semantics Guided Data Selection (DSGDS) strategy is further introduced to ensure that synthesized samples contribute to balanced and discriminative feature distributions within each domain. Building on these components, a collaborative multi-domain classification scheme is developed to jointly leverage domain-shared and domain-specific representations, adaptively integrating their complementary information through an uncertainty-aware fusion mechanism at the decision level. As a result, IMDG robustly produces reliable classification predictions on unseen target data. Experimental evaluations on heterogeneous datasets show that IMDG outperforms existing state-of-the-art methods in this challenging and practical setting.