Systematic examination of the incorporation of class relationships via multilabel, multiclass, hierarchical classification

Abstract

Difficulties in multiclass classification problems often arise as a result of the transformation to a series of one‐class problems, where each class is determined without respect to the results from the other classes. Existing relationships between the classes themselves, rather than the individual samples in each class, are ignored, and the information is not incorporated into the classification model. Alternatively, classifying multiclass samples via a multilabel, hierarchical classification method can incorporate this information through alternative means of sample selection and comparison. In this paper, the possible methods are systematically examined and the results ranked in terms of two metrics, accuracy and F1 measure. This method constructed several models that outperformed models constructed by traditional classification methods, showing that incorporating the additional information available in class relationships offers an advantage in performance over traditional “flat” classification.

Index, Volume 3, 2022

“…The advantage of a flat multiclass classifier lies in its simplicity and ease of implementation. The notable disadvantage of this model is that it is possible to lose valuable information that may lie in the natural hierarchy of the data by ignoring parent-child class relationships (Daisey and Brown, 2017).…”