Error rate control for classification rules in multiclass mixture models.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Finite mixture models are widely used for data clustering and classification.
• Controlling classification error rates is crucial for reliable model interpretation.
• Existing methods often use thresholded Maximum A Posteriori (MAP) rules, which can be conservative.

Purpose of the Study:

• To develop an optimal classification rule for finite mixture models that minimizes Type II error rates while controlling Type I error rates.
• To define and evaluate a multiclass False Discovery Rate (FDR)-like rule.
• To compare the performance of the proposed rule against standard thresholded MAP rules.

Main Methods:

• Defining Type I and Type II classification error rates analogous to statistical test theory.
• Identifying an optimal region in the observation space for applying the MAP rule.
• Developing a heuristic for computing the optimal classification rule.
• Implementing and comparing a multiclass FDR-like rule with thresholded MAP rules.

Main Results:

• An optimal classification rule is found to correspond to searching an optimal region in the observation space.
• The shape of this optimal region depends on the misclassification rate to be controlled.
• The proposed multiclass FDR-like optimal rule demonstrates a less conservative approach compared to thresholded MAP rules.
• Validation on simulated and real datasets confirms the practical advantages of the FDR-like rule.

Conclusions:

• The developed optimal classification rule offers improved performance in finite mixture models.
• The multiclass FDR-like rule provides a more powerful alternative to conventional thresholded MAP rules.
• This approach enhances the ability to classify observations accurately while maintaining control over error rates.