Unobserved classes and extra variables in high-dimensional discriminant analysis.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Supervised classification models face challenges when test data includes unobserved classes or additional variables not present during training.
• Existing classifiers struggle to adapt to evolving datasets with new classes and increased dimensionality.

Purpose of the Study:

• To introduce a novel discriminant approach, Dimension-Adaptive Mixture Discriminant Analysis (D-AMDA), capable of handling unobserved classes and adapting to growing data dimensions.
• To develop a robust framework for adaptive variable selection and classification in high-dimensional datasets.

Main Methods:

• Developed Dimension-Adaptive Mixture Discriminant Analysis (D-AMDA), a model-based discriminant approach.
• Employed an Expectation-Maximization (EM) algorithm for model estimation via a full inductive approach.
• Integrated D-AMDA into a general framework for adaptive variable selection and classification.

Main Results:

• D-AMDA demonstrated the ability to detect unobserved classes in test data.
• The proposed framework successfully adapted to increasing data dimensionality.
• Validation through simulation and an adulterated honey classification experiment confirmed the method's efficacy in complex scenarios.

Conclusions:

• D-AMDA provides an effective solution for supervised classification problems with evolving data characteristics.
• The adaptive framework enhances classifier performance when dealing with unknown classes and high-dimensional data.
• The approach is suitable for real-world applications requiring robust and adaptive classification models.