Unsupervised Bayesian classification for models with scalar and functional covariates.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Traditional classification methods struggle with high-dimensional functional data.
• Existing models often fail to capture the inherent structure of complex datasets.
• The curse of dimensionality is a significant challenge in analyzing functional covariates.

Purpose of the Study:

• To develop a flexible unsupervised classification framework for mixture models.
• To incorporate both scalar and functional covariates effectively.
• To address limitations of existing methods in handling complex data structures.

Main Methods:

• A hierarchical Bayesian model with a latent multinomial variable is proposed.
• Basis expansions are utilized to reduce the dimensionality of functional covariates.
• A generalized linear model links mixture probabilities to covariates.

Main Results:

• The proposed method provides accurate parameter estimation and latent classification predictions.
• Demonstrated effectiveness on real-world examples, including clinical trial response identification and disease prediction in livestock.
• Successfully overcomes the curse of dimensionality associated with functional data.

Conclusions:

• The novel Bayesian approach offers a robust solution for unsupervised classification with mixed data types.
• The method enhances predictive accuracy and interpretability in complex scenarios.
• Applicable to diverse fields requiring sophisticated data analysis, such as medicine and agriculture.