An Empirical Comparison of Flat and Hierarchical Performance Measures for Multi-Label Classification with Hierarchy Extraction

Abstract

Multi-label Classification (MC) often deals with hierarchically organized class taxonomies. In contrast to Hierarchical Multi-label Classification (HMC), where the class hierarchy is assumed to be known a priori, we are interested in the opposite case where it is unknown and should be extracted from multi-label data automatically. In this case the predictive performance of a classifier can be assessed by well-known Performance Measures (PMs) used in flat MC such as precision and recall. The fact that these PMs treat all class labels as independent labels, in contrast to hierarchically structured taxonomies, is a problem. As an alternative, special hierarchical PMs can be used that utilize hierarchy knowledge and apply this knowledge to the extracted hierarchy. This type of hierarchical PM has only recently been mentioned in literature. The aim of this study is first to verify whether HMC measures do significantly improve quality assessment in this setting. In addition, we seek to find a proper measure that reflects the potential quality of extracted hierarchies in the best possible way. We empirically compare ten hierarchical and four traditional flat PMs in order to investigate relations between them. The performance measurements obtained for predictions of four multi-label classifiers ML-ARAM, ML-kNN, BoosTexter and SVM on four datasets from the text mining domain are analyzed by means of hierarchical clustering and by calculating pairwise statistical consistency and discriminancy.