Inconsistent Node Flattening for Improving Top-Down Hierarchical Classification

摘要: Large-scale classification of data where classes are structurally organized in a hierarchy is an important area research. Top-down approaches that exploit the during learning and prediction phase efficient for large-scale hierarchical classification. However, accuracy top-down poor due to error propagation i.e., errors made at higher levels cannot be corrected lower levels. One main reason behind presence inconsistent nodes introduced arbitrary process creating these hierarchies by domain experts. In this paper, we propose two different data-driven (local global) structure modification identifies flattens present within hierarchy. Our extensive empirical evaluation proposed on several image text datasets with varying distribution features, training instances per class shows improved performance over competing approaches. Specifically, see improvement upto 7% Macro-F1 score our approach best TD baseline. SOURCE CODE: http://www.cs.gmu.edu/ mlbio/InconsistentNodeFlattening.