Avoiding inferior clusterings with misspecified Gaussian mixture models.

Area of Science:

• Data Science
• Statistical Modeling
• Machine Learning

Background:

• Clustering is vital for exploratory data analysis across scientific fields.
• Gaussian Mixture Models (GMMs) are popular for clustering but can fail with non-Gaussian or noisy data.
• Misspecified GMMs can lead to incorrect classifications and inferior clustering solutions.

Purpose of the Study:

• To identify and characterize inferior clustering solutions arising from misspecified GMMs.
• To develop a novel GMM-based clustering algorithm that avoids both inferior and spurious solutions.
• To improve the robustness and interpretability of clustering when underlying data distributions are unknown.

Main Methods:

• Theoretical analysis of component asymmetry in misspecified GMMs.
• Introduction of a new penalty term to mitigate inferior and spurious solutions.
• Development of a novel clustering algorithm, SIA, incorporating a new model selection criterion.

Main Results:

• Characterization of inferior clustering solutions, distinct from spurious solutions.
• Demonstration that the proposed penalty term effectively avoids inferior solutions.
• Empirical evidence showing SIA outperforms existing GMM-based methods in misspecified scenarios.

Conclusions:

• The proposed SIA algorithm offers a robust solution for clustering with potentially misspecified GMMs.
• SIA enhances the reliability of clustering by avoiding problematic solution types.
• This work contributes to more accurate data analysis in diverse scientific applications.