Systematic tuning of parameters in support vector clustering.

Area of Science:

• Data Science
• Machine Learning
• Computational Statistics

Background:

• Clustering algorithms group data based on shared features.
• Arbitrary parameter tuning often leads to suboptimal clustering results.
• Efficient and systematic parameter optimization is crucial for effective clustering.

Purpose of the Study:

• To present a global optimization strategy for systematic and optimal selection of clustering algorithm parameters.
• To propose and benchmark a novel performance criterion against existing metrics.
• To introduce a computationally efficient contour plotting approach for cluster labeling.

Main Methods:

• Global optimization strategy using simulated annealing for parameter tuning.
• Development and benchmarking of a new performance criterion.
• Implementation of a contour plotting approach as an alternative to adjacency matrices for cluster assignment.
• Validation on diverse datasets including UCI (iris, thyroid) and medical data (lymphoma, breast cancer).

Main Results:

• The proposed strategy efficiently determines optimal tuning parameters for clustering algorithms like Support Vector Clustering (SVC).
• The contour plotting approach significantly reduces computational time (CPU time), especially for large datasets.
• The proposed performance criterion showed competitive or superior performance compared to Silhouette, Dunn's, and Davies-Bouldin indices, with mixed results for existing metrics.

Conclusions:

• The developed global optimization strategy offers an efficient method for optimal parameter selection in clustering.
• The contour plotting approach provides a computationally advantageous alternative for cluster labeling.
• The proposed performance criterion demonstrates potential for evaluating clustering quality, outperforming Dunn's index and showing comparable results to Silhouette and Davies-Bouldin.