Hierarchical confusion matrix for classification performance evaluation
Kevin Riehl, Michael Neunteufel, Martin Hemberg
Abstract
Abstract This study proposes the novel concept of hierarchical confusion matrix, opening the door for popular confusion-matrix-based (flat) evaluation measures from binary classification problems, while considering the peculiarities of hierarchical classification problems. The concept is developed to a generalised form and proven its applicability to all types of hierarchical classification problems including directed acyclic graphs, multi-path labelling, and non-mandatory leaf-node prediction. Finally, measures based on the novel confusion matrix are used for three real-world hierarchical classification applications and compared to established evaluation measures. The results, the conformity with important attributes of hierarchical classification schemes and its broad applicability justify its recommendation.