A Comparison Study of Validity Indices on Swarm-Intelligence-Based Clustering.

Area of Science:

• Computational intelligence
• Data mining
• Machine learning

Background:

• Swarm intelligence offers advantages for clustering, including parallel processing and avoiding local minima.
• Clustering validity indices serve as fitness functions to assess cluster quality in swarm intelligence.
• The performance of swarm intelligence clustering is sensitive to the choice of validity index, as they are data-dependent.

Purpose of the Study:

• To compare the performance of eight common clustering validity indices.
• To evaluate these indices within a differential-evolution-particle-swarm-optimization (DEPSO) clustering framework.
• To identify the most effective validity index for swarm intelligence-based clustering.

Main Methods:

• Implemented differential-evolution-particle-swarm-optimization (DEPSO), a hybrid algorithm combining differential evolution and particle swarm optimization.
• Applied eight distinct clustering validity indices: Caliński-Harabasz, CS, Davies-Bouldin, Dunn (and two generalized versions), I, and silhouette statistic.
• Tested the indices on both synthetic and real-world datasets.

Main Results:

• The silhouette statistic index demonstrated superior performance across a majority of the tested datasets.
• Differential-evolution-particle-swarm-optimization (DEPSO) enhanced search capabilities and flexibility in exploring the problem space.
• Variations in index selection significantly impacted the quality of the obtained clusters.

Conclusions:

• The silhouette statistic index is a highly effective validity index for swarm intelligence-based clustering.
• Relying on a single index may be insufficient; considering multiple indices is recommended for robust clustering.
• The choice of validity index critically influences the reliability of clustering structures derived from swarm intelligence methods.