Normalized Hierarchical SVM

TL;DR

This paper introduces normalized hierarchical SVMs that improve large-scale hierarchical classification by normalizing regularization and margin, achieving state-of-the-art results and connecting to matrix factorization models.

Contribution

It proposes normalization techniques for structured SVMs in hierarchical classification, significantly enhancing performance and establishing a link to matrix factorization.

Findings

Normalized hierarchical SVMs outperform unnormalized models.

Normalization enables state-of-the-art results in large-scale hierarchical tasks.

The approach reveals a connection between hierarchical SVMs and matrix factorization.

Abstract

We present improved methods of using structured SVMs in a large-scale hierarchical classification problem, that is when labels are leaves, or sets of leaves, in a tree or a DAG. We examine the need to normalize both the regularization and the margin and show how doing so significantly improves performance, including allowing achieving state-of-the-art results where unnormalized structured SVMs do not perform better than flat models. We also describe a further extension of hierarchical SVMs that highlight the connection between hierarchical SVMs and matrix factorization models.