Clustering algorithms: A comparative approach.

Area of Science:

• Data Science
• Machine Learning
• Pattern Recognition

Background:

• Machine learning methods are crucial for real-world pattern recognition tasks.
• No consensus exists on the most suitable classification methods for diverse datasets.
• Systematic comparison of clustering algorithms is essential for practical applications.

Purpose of the Study:

• To systematically compare nine well-known clustering methods in R.
• To evaluate method performance on artificial datasets with varying properties.
• To assess the sensitivity of clustering methods to parameter configurations.

Main Methods:

• Comparison of 9 clustering algorithms in the R programming language.
• Generation of artificial datasets with tunable properties (e.g., number of classes, class separation).
• Evaluation of method sensitivity to parameter settings and default configurations.

Main Results:

• Spectral clustering demonstrated strong performance with default settings across various scenarios.
• Default configurations of some clustering methods were not consistently accurate.
• Random parameter selection was an effective strategy to enhance clustering performance when defaults failed.

Conclusions:

• The spectral clustering approach is a robust choice for normally distributed data.
• Parameter tuning is critical for optimizing clustering algorithm performance.
• Guidance is provided for selecting appropriate clustering algorithms based on data characteristics and performance.