Handling skewness and directional tails in model-based clustering

Area of Science:

• Statistics
• Data Mining
• Machine Learning

Background:

• Model-based clustering is vital for identifying data patterns.
• Standard methods struggle with skewed or heavy-tailed data clusters.
• Outlier detection remains a challenge in complex datasets.

Purpose of the Study:

• Introduce novel clustering models using transformed MSCN distributions.
• Enhance cluster shape flexibility (skewness, kurtosis).
• Enable component-wise and directional outlier detection.

Main Methods:

• Developed two models based on component-wise transformations of observed data.
• Utilized mixture of multiple scaled contaminated normal (MSCN) distributions.
• Incorporated directional outlier detection using principal components.

Main Results:

• Proposed MSCN-based clustering offers flexible cluster shapes.
• Achieved effective component-wise and directional outlier detection.
• Demonstrated advantages over global/component-wise outlier detection methods.

Conclusions:

• The MSCN-based approach provides robust clustering for complex data.
• Offers superior outlier detection capabilities in various dimensions.
• Outperforms existing methods in specific practical clustering scenarios.