Regularization Framework for Large Scale Hierarchical Classification

In this paper, we propose a hierarchical regularization framework for large-scale hierarchical classification. In our framework, we use the regularization structure to share information across the hierarchy and enforce similarity between class-parameters that are located nearby in the hierarchy. To address the computational issues that arise, we propose a parallel-iterative optimization scheme that can handle large-scale problems with tens of thousands of classes and hundreds of thousands of instances. Experiments on multiple benchmark datasets showed significant performance improvements of our proposed approach over other competing approaches.