Cross-Scale Partial Order Metric Learning for Hierarchical Label Structures

Multi-scale label hierarchies with diverse semantic levels often give rise to similarity conflicts among sample pairs, posing a critical challenge to conventional deep metric learning (DML) methods. To address this issue, we propose Cross-Scale Partial Order Metric Learning (CSPOML), which is designed to enhance the hierarchical consistency of embeddings across different semantic levels. Specifically, we introduce a partial-order weight function to quantify the consistency of higher-level ancestor labels, and we design a cross-scale positive-pair consistency constraint together with a negative-pair trend-suppression mechanism to regulate similarity behavior across scales. Furthermore, we incorporate an uncertainty-aware proxy modeling strategy at the fine level to alleviate the noise introduced by ambiguous samples, thereby improving the structural alignment of the embedding space. Extensive experiments on three dynamic hierarchical metric learning datasets, DyML-Vehicle, DyML-Animal, and DyML-Product, show that CSPOML achieves consistent improvements on mAP and ASI while remaining competitive on Recall@1 across different semantic levels. These results support the effectiveness of the proposed cross-scale partial-order modeling.