Label Hierarchy Transition: Delving Into Class Hierarchies to Enhance Deep Classifiers.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Hierarchical classification organizes objects into nested categories.
• Current methods often treat levels independently, missing inter-level category correlations.
• This limitation hinders performance in complex classification tasks.

Purpose of the Study:

• To propose a unified deep learning framework, Label Hierarchy Transition (LHT), for effective hierarchical classification.
• To address the limitations of existing multi-task learning strategies in exploiting category correlations.
• To enhance the accuracy and robustness of hierarchical classification models.

Main Methods:

• Developed a unified probabilistic framework, Label Hierarchy Transition (LHT).
• Incorporated a transition network to learn label hierarchy transition matrices.
• Introduced a confusion loss to enforce learning of cross-hierarchy correlations.
• The framework is adaptable to existing deep networks with minimal changes.

Main Results:

• LHT framework demonstrated superior performance on public benchmark datasets for hierarchical classification.
• The transition network effectively encoded underlying correlations within class hierarchies.
• The confusion loss facilitated learning of correlations across different label hierarchies.
• Achieved state-of-the-art results, surpassing existing methods.

Conclusions:

• LHT offers a superior approach to hierarchical classification by explicitly modeling label transitions.
• The framework shows significant potential for applications like computer-aided diagnosis, exemplified by skin lesion diagnosis.
• LHT provides a robust and adaptable solution for complex hierarchical classification problems.