Optimizing the C Index Using a Canonical Genetic Algorithm

Clustering is an important family of unsupervised machine learning methods. Cluster validity indices are widely used to assess the quality of obtained clustering results. The C index is one of the most popular cluster validity indices. This paper shows that the C index can be used not only to validate but also to actually find clusters. This leads to difficult discrete optimization problems which can be approximately solved by a canonical genetic algorithm. Numerical experiments compare this novel approach to the well-known c-means and single linkage clustering algorithms. For all five considered popular real-world benchmark data sets the proposed method yields a better C index than any of the other (pure) clustering methods.