Joint Learning of Hyperbolic Label Embeddings for Hierarchical Multi-label Classification

We consider the problem of multi-label classification where the labels lie in\na hierarchy. However, unlike most existing works in hierarchical multi-label\nclassification, we do not assume that the label-hierarchy is known. Encouraged\nby the recent success of hyperbolic embeddings in capturing hierarchical\nrelations, we propose to jointly learn the classifier parameters as well as the\nlabel embeddings. Such a joint learning is expected to provide a twofold\nadvantage: i) the classifier generalizes better as it leverages the prior\nknowledge of existence of a hierarchy over the labels, and ii) in addition to\nthe label co-occurrence information, the label-embedding may benefit from the\nmanifold structure of the input datapoints, leading to embeddings that are more\nfaithful to the label hierarchy. We propose a novel formulation for the joint\nlearning and empirically evaluate its efficacy. The results show that the joint\nlearning improves over the baseline that employs label co-occurrence based\npre-trained hyperbolic embeddings. Moreover, the proposed classifiers achieve\nstate-of-the-art generalization on standard benchmarks. We also present\nevaluation of the hyperbolic embeddings obtained by joint learning and show\nthat they represent the hierarchy more accurately than the other alternatives.\n