Sparse Feature-Persistent Hierarchical Classification

Hierarchical methods that divide target classes into nested sets of progressive generality are increasingly important in large-scale classification tasks. Additionally, large-scale classification over many classes typically requires a large number of features, which necessitates larger amounts of training data. In this work, we leverage class hierarchies to mitigate data paucity in large-scale problems. We combine sparse classification methods with a hidden Markov tree model to identify and exploit feature saliency across different levels in a class taxonomy. By modeling the hierarchical persistence of salient features, the proposed method is designed to improve classification performance in scenarios where training data is limited and high dimensional. Examples demonstrate efficacy of the approach on several measured datasets.