A robustness metric for biological data clustering algorithms.

Area of Science:

• Data Science
• Bioinformatics
• Computational Biology

Background:

• Cluster analysis is fundamental in data-centric computation, with algorithm selection influenced by data characteristics and desired cluster types.
• Comparing clustering algorithms often focuses on cluster quality, but algorithm settings significantly impact results.
• The variability of clustering outputs based on setting changes is a critical, yet often overlooked, aspect.

Purpose of the Study:

• To introduce a novel metric, "robustness," for quantifying the stability of clustering algorithm outputs.
• To evaluate the robustness of eleven popular clustering algorithms across diverse datasets.
• To provide a tool for informed algorithm selection and parameter tuning.

Main Methods:

• Development of the "robustness" metric to assess output consistency across algorithm settings.
• Evaluation of eleven clustering algorithms on approximately two dozen publicly available mRNA expression microarray datasets.
• Comparative analysis of algorithm performance based on the calculated robustness scores.

Main Results:

• Hierarchical clustering methods generally demonstrated the highest robustness due to their straightforward and predictable nature.
• The paraclique algorithm exhibited consistently high robustness, often matching or exceeding hierarchical methods on several datasets.
• Other tested algorithms showed variable robustness, with no clear performance distinctions among them.

Conclusions:

• Robustness offers a simple, interpretable measure of clustering algorithm stability and predictability.
• This metric can assist researchers in selecting appropriate algorithms and optimizing parameter tuning efforts.
• Understanding algorithm robustness is crucial for reliable data analysis in fields like bioinformatics.