Learning Ordinal-Hierarchical Constraints for Deep Learning Classifiers.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Real-world classification often involves categories with inherent hierarchical and ordinal relationships.
• Existing deep learning (DL) models struggle to simultaneously capture these hierarchical and ordinal constraints, limiting generalization performance.

Purpose of the Study:

• To propose novel DL methodologies that effectively model both hierarchical and ordinal structures in classification.
• To enhance generalization performance in complex classification problems by integrating these constraints.

Main Methods:

• Introduced two new ordinal-hierarchical DL methodologies: hierarchical cumulative link model (HCLM) and hierarchical-ordinal binary decomposition (HOBD).
• Decomposed the problem into local and global graph paths to encode ordinal constraints at each hierarchical level.
• Framed the problem as minimizing combined global and local losses, utilizing ordinal binary decomposition (OBD) and cumulative link model (CLM) for constraint setting.

Main Results:

• The proposed HCLM and HOBD models demonstrated statistically significant improvements over state-of-the-art methods.
• Effectiveness was validated across diverse real-world datasets from industrial, biomedical, computer vision, and financial domains.
• The models successfully captured and leveraged both hierarchical and ordinal label structures.

Conclusions:

• The novel ordinal-hierarchical DL approaches offer a significant advancement for classification tasks with complex label structures.
• These methodologies provide a robust framework for improving generalization in real-world applications across various scientific and industrial fields.