Detecting influential observations in a model-based cluster analysis.

Area of Science:

• Statistics
• Data Science

Background:

• Finite mixture models are valuable for analyzing population heterogeneity and complex data like repeated-measurements.
• Model-based clustering performance is significantly affected by outliers and influential data points.
• Existing methods primarily focus on outlier detection, with fewer approaches for identifying influential observations.

Purpose of the Study:

• To develop and apply local-influence diagnostics for finite mixture models.
• To address the challenge of identifying influential observations in mixture modeling.
• To improve the robustness of clustering and classification using finite mixture models.

Main Methods:

• Application of local-influence diagnostics tailored to finite mixture models.
• Focus on models with a known number of components.
• Illustration using real-life datasets to demonstrate practical utility.

Main Results:

• Local-influence diagnostics effectively identify influential observations in finite mixture models.
• The proposed methodology provides a practical tool for data analysis.
• Enhanced understanding of data structure and model sensitivity.

Conclusions:

• Local-influence diagnostics are a valuable addition to the toolkit for finite mixture modeling.
• The approach enhances the interpretability and reliability of clustering and classification results.
• Further research can explore extensions to models with an unknown number of components.