Exploitation of pairwise class distances for ordinal classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Ordinal classification problems involve ordered categories, posing unique challenges for predictive modeling.
• Existing methods often train latent space projections indirectly, limiting model interpretability and quality.

Purpose of the Study:

• To develop a direct projection model for ordinal classification using insights from pairwise distance calculations.
• To enhance the quality and understandability of latent models in ordinal classification tasks.

Main Methods:

• A novel methodology constructs a direct projection model by analyzing class distribution through pairwise distances.
• The approach was evaluated against 8 established classification methods on 10 real-world datasets using 4 performance metrics.

Main Results:

• The proposed method achieved superior average ranking across three out of four performance metrics.
• While competitive, significant differences were only observed for specific comparison methods.
• Analysis revealed that existing methods' latent space projections do not fully capture intraclass pattern behavior.

Conclusions:

• The new direct projection method offers a simple, intuitive, and highly competitive alternative for ordinal classification.
• This approach provides a more direct and understandable way to model the latent space in ordinal classification problems.