Robust representation learning from multi-source data necessitates the effective orchestration of complementary information while preserving semantic integrity. Existing methods primarily focus on class-level or instance-level alignment, neglecting fine-grained feature consistency and hierarchical collaborative mechanisms, which consequently limits representation precision. To address these issues, we propose the Tri-Level Consistency–Diversity Calibration (TCDC) method, a hierarchical framework designed to optimize information flow across feature, instance, and class levels. Specifically, at the feature level, TCDC imposes a variance–covariance constraint to align fine-grained features, thereby decorrelating dimensions. At the instance level, semantics-guided multi-objective graph learning is integrated with contrastive learning to adaptively calibrate global topology and capture high-order category correlations. Finally, a class-level attraction–repulsion constraint leverages category prototypes as global semantic anchors to promote intra-class aggregation and enhance inter-class separability. Extensive experiments on multiple public datasets demonstrate the effectiveness of TCDC.
Hu et al. (Mon,) studied this question.