Outlier Identification in Model-Based Cluster Analysis.

Area of Science:

• Statistics
• Data Mining
• Machine Learning

Background:

• Model-based clustering using normal-mixture models can be sensitive to outlying observations, potentially distorting cluster structure and count.
• Existing methods may struggle with accurate outlier identification in complex datasets.

Purpose of the Study:

• To develop and validate a robust method for identifying outlier observations within normal-mixture model-based clusters.
• To enhance the reliability of clustering results by mitigating the influence of outliers.

Main Methods:

• The proposed method identifies outliers based on minimal cluster membership proportion or significant changes in cluster-specific variance.
• Utilizes the MCLUST R package with a modified prior for cluster-specific variance to prevent estimation degeneracies.
• Evaluated through simulation studies and comparison with existing outlier detection techniques.

Main Results:

• The developed method demonstrates high accuracy in detecting true outliers while minimizing false positives.
• Outperforms other approaches in outlier detection across various simulated scenarios.
• Applied to National Hockey League data, showing comparable results to published findings.

Conclusions:

• The novel outlier identification method significantly improves the robustness of normal-mixture model-based clustering.
• Offers a valuable tool for data analysis where outlier detection is critical for accurate cluster interpretation.
• The modified prior for variance estimation addresses key challenges in model-based clustering procedures.