Infinite Bayesian Max-Margin Discriminant Projection.

Area of Science:

• Machine Learning
• Data Science
• Pattern Recognition

Background:

• Supervised dimensionality reduction is crucial for complex datasets.
• Model selection, such as determining the number of local regions, presents a challenge.
• Handling nonlinear data structures requires advanced techniques.

Purpose of the Study:

• To propose an infinite Bayesian max-margin linear discriminant projection (iMMLDP) model.
• To extend the iMMLDP model to a nonlinear version (iKMMDP) using the kernel trick.
• To integrate dimensionality reduction, clustering, and classification within a unified framework.

Main Methods:

• Developed iMMLDP using Bayesian nonparametric priors for automatic model selection.
• Formulated iKMMDP to address local nonlinear separable structures via kernel methods.
• Employed Gibbs sampling for efficient parameter inference due to conjugate properties.

Main Results:

• Demonstrated the effectiveness of iMMLDP and iKMMDP on synthesized and real-world data.
• Validated the models on multimodally distributed datasets and radar image data.
• Showcased the models' ability to combine dimensionality reduction, clustering, and classification.

Conclusions:

• The proposed iMMLDP and iKMMDP models offer a principled approach to supervised dimensionality reduction.
• These Bayesian methods efficiently handle model selection and complex data structures.
• The models are effective for analyzing diverse datasets, including challenging real-world applications.