Ordinal regression by a generalized force-based model.

Area of Science:

• Machine Learning
• Computer Science
• Statistics

Background:

• Multiclass classification problems with ordered classes present unique challenges.
• Existing gravitational models offer a framework but require adaptation for ordinal data.
• Interpretability is crucial for real-world applications of classification algorithms.

Purpose of the Study:

• To introduce a new instance-based algorithm for multiclass classification with natural class order.
• To extend state-of-the-art gravitational models by generalizing class-pattern interaction forces.
• To improve interpretability and generalization performance in ordinal regression.

Main Methods:

• The algorithm generalizes gravitational models by modifying class-pattern interaction forces.
• It incorporates a weight matrix for attribute space metric modification and class-specific force law parameters.
• A probabilistic error function enables parameter estimation via optimization, penalizing non-unimodal outputs.

Main Results:

• The proposed algorithm demonstrates competitive generalization performance on ordinal regression datasets.
• Experimental results on both discretized and real-world ordinal datasets validate the model's effectiveness.
• Nonparametric statistical tests confirm the algorithm's performance compared to existing methods.

Conclusions:

• The novel algorithm effectively addresses multiclass classification problems with ordered classes.
• Its generalized gravitational approach and interpretability make it suitable for practical applications.
• The method offers a promising advancement in ordinal regression techniques.