Output details

Finding reproducible cluster partitions for the k-means algorithm

<13> Cluster analysis. The significance of the paper is to define performance criteria to ensure repeatability of k-means clustering while maintaining good cluster separation. The originality is to propose and validate a systematic methodology to measure reproducibility of cluster partitions in a rigorous statistical manner and show that this is a complementary performance criterion to the usual quadratic score. Combining the two criteria is effective for selecting the most reproducible cluster partitions from the usual multiple random initialisations. This can be automated and will remove the level of arbitrariness that is prevalent in many practical applications of k-means clustering.