Tackling Ordinal Regression Problem for Heterogeneous Data: Sparse and Deep Multi-Task Learning Approaches.

Area of Science:

• Machine Learning
• Data Science
• Biomedical Informatics

Background:

• Real-world data often features ordinal labels, requiring specialized regression techniques.
• Existing ordinal regression methods typically assume independent and identically distributed (IID) data, limiting their application to heterogeneous datasets.
• Multi-task learning (MTL) offers a framework for learning from related but distinct tasks, showing promise for non-IID data.

Purpose of the Study:

• To address the gap in applying multi-task learning (MTL) to ordinal regression for heterogeneous, non-IID data.
• To develop novel Multi-Task Ordinal Regression (MTOR) models capable of handling complex data structures.
• To improve prediction performance in domains with ordered categorical outcomes.

Main Methods:

• Developed a regularized MTOR model for smaller datasets.
• Developed a deep neural network-based MTOR model for large-scale datasets.
• Evaluated models on three real-world healthcare datasets for multi-stage disease progression diagnosis.

Main Results:

• The proposed MTOR models demonstrated a marked improvement in prediction performance.
• MTOR models outperformed traditional single-task ordinal regression approaches.
• Effectiveness shown in diagnosing multi-stage disease progression.

Conclusions:

• MTOR provides a powerful framework for ordinal regression on heterogeneous, non-IID data.
• The developed MTOR models offer enhanced predictive accuracy in complex, real-world scenarios.
• This approach has significant implications for data analysis in health sciences and other fields utilizing ordinal data.