Suhyeon Jo, Hierarchical Multi-Label Classification from Partial Labels without Known Hierarchy, Master's Thesis, Department of Industrial and Systems Engineering, KAIST, 2023 

Abstract

Multi-label classification aims to find a function that predicts multiple labels for an instance. For successful learning, the classifier requires a fully ground-truth label dataset which demands a high annotation cost. Many previous works proposed classification model learning from partial labels to alleviate the issue. However, those works have limitations due to the assumption that classes are mutually exclusive. In a realistic setting, classes are not disjoint, i.e., a hierarchical relationship exists between categories. When classes of a dataset are organized into a hierarchical structure, Hierarchical Multi-label Classification (HMC) aims to learn a multi-label classifier that satisfies hierarchical constraints. Previous works impose hierarchical constraints on network architectures or loss functions to predict an instance with a set of hierarchically related labels. To the best of our knowledge, we propose for the first time the problem to find a hierarchical multi-label classifier when a partial label dataset without a known hierarchy is given rather than a fully-supervised dataset. In this work, we propose an iterative framework for hierarchical multi-label classification from partial labels without known hierarchy. When training a multi-label classifier with partial labels, our model extracts label hierarchy from the classifier output using our hierarchy extraction algorithm. Then, our proposed losses exploit the extracted hierarchy to train the classifier. Several experiments show that our model obtains a label hierarchy quite close to the ground-truth dataset hierarchy, so that outperforms previous methods for multi-label classification from partial labels.

@masterthesis{Jo:2023,

author = {Suhyeon Jo},

advisor ={Il-Chul Moon},

title = {Hierarchical Multi-Label Classification from Partial Labels without Known Hierarchy},

school = {KAIST},

year = {2023}

}