Hierarchical metric learning with intra-level and inter-level regularization

Metric learning for hierarchical classification is a significant problem whose purpose is to learn more discriminative metrics by exploiting the dataset's hierarchical structure and achieving higher accuracy rates for hierarchical classification. However, most of the existing hierarchical metric learning methods fail to consider the irrelevance between the metrics of sibling nodes in a hierarchical tree, which makes the metric of each node not well distinguish child nodes. This paper proposes a hierarchical metric learning model based on intra-level and inter-level regularization. The model mines the idiosyncrasies of sibling nodes and learns a more discriminative metric for each non-leaf node. At the same time, by exploiting the commonalities of parent-child nodes to control inter-level error propagation. Extensive experiments on five hierarchical datasets demonstrate that the proposed algorithm can perform better than the existing ones.