Clustering benchmark datasets exploiting the fundamental clustering problems.

Area of Science:

• Data Science
• Machine Learning
• Computer Science

Background:

• Clustering algorithms are essential for data analysis, but their performance varies across different data structures.
• Existing benchmarks may not fully capture the complexities of real-world, high-dimensional datasets.
• The Fundamental Clustering Problems Suite (FCPS) was developed to address these limitations.

Purpose of the Study:

• To introduce the Fundamental Clustering Problems Suite (FCPS) as a comprehensive benchmark for clustering algorithms.
• To provide datasets designed to challenge and evaluate the capabilities of clustering and dimensionality reduction methods.
• To facilitate the investigation of algorithm shortcomings, particularly in higher dimensions.

Main Methods:

• The FCPS comprises datasets with known classifications, designed for visualization in 2D or 3D.
• Datasets are intentionally crafted to represent specific clustering challenges.
• Includes user-defined sample sizes via an R package and distance matrices for high-dimensional datasets (Leukemia, Tetragonula).

Main Results:

• The FCPS datasets highlight varying success rates of known clustering algorithms.
• Demonstrates the utility of FCPS in revealing limitations of dimensionality reduction techniques for datasets beyond 3D.
• Provides a standardized suite for comparative analysis of clustering algorithm performance.

Conclusions:

• The FCPS is a valuable resource for assessing the robustness and limitations of clustering algorithms.
• It serves as a critical tool for advancing the development of more effective clustering and dimensionality reduction methods.
• The suite is particularly relevant for tackling challenges posed by high-dimensional and complex data structures.