Mixture-Based Probabilistic Graphical Models for the Label Ranking Problem.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Probabilistic Graphical Models

Background:

• The Label Ranking (LR) problem requires learning preference models for unlabeled instances.
• Existing methods like k-nearest neighbors and decision trees perform well.
• Probabilistic Graphical Models (PGMs) have been underexplored due to challenges in modeling permutations.

Purpose of the Study:

• To propose a novel Hidden Naive Bayes classifier (HNB) for addressing the LR problem.
• To integrate diverse probability distributions (multinomial, Gaussian, Mallows) within a hybrid Bayesian network framework.
• To explore two PGM approaches: Naive Bayes structure and models with attribute interactions.

Main Methods:

• Developed a Hidden Naive Bayes classifier (HNB) incorporating a hidden variable.
• Designed a hybrid Bayesian network capable of combining multinomial, Gaussian, and Mallows distributions.
• Investigated two probabilistic models: one with a Naive Bayes structure and another with multivariate Gaussian distributions for parameter estimation.

Main Results:

• The proposed HNB classifier demonstrates competitive performance against state-of-the-art algorithms.
• The models show comparable accuracy in predicting preferred label rankings.
• Experimental evaluations indicate efficient CPU time requirements.

Conclusions:

• The Hidden Naive Bayes classifier offers a viable and effective solution for the Label Ranking problem.
• Hybrid Bayesian networks can successfully model permutations using distributions like Mallows.
• The proposed methods present a promising alternative to existing LR algorithms.