Diluvian Clustering: A Fast, Effective Algorithm for Clustering Compositional and Other Data.

Area of Science:

• Computational Statistics
• Data Mining
• Machine Learning

Background:

• Traditional clustering algorithms often struggle with noisy and compositional datasets.
• Existing methods for compositional data typically rely on specific metrics, limiting their applicability.
• The need for robust algorithms capable of handling clusters with varying characteristics (size, density) is critical.

Purpose of the Study:

• Introduce Diluvian Clustering, a new unsupervised grid-based clustering algorithm.
• Demonstrate its effectiveness in interpreting large, noisy compositional datasets.
• Highlight its advantages over metric-dependent clustering approaches.

Main Methods:

• Developed a grid-based clustering algorithm that does not depend on an explicit metric.
• Implemented a method capable of identifying both compact and diffuse clusters of varying sizes.
• Ensured computational efficiency, with typical performance near O(N).

Main Results:

• Diluvian Clustering successfully clusters large datasets (e.g., 20,000 particles, 30 elements) rapidly.
• The algorithm effectively handles clusters with disparate variances, common in real-world data.
• Identified clusters with diverse properties, including variations in size and density.

Conclusions:

• Diluvian Clustering offers a computationally efficient and flexible alternative for compositional data analysis.
• Its metric-independent nature allows for robust clustering of diverse real-world datasets.
• The algorithm's intuitive parameters and performance make it suitable for large-scale data interpretation.