Comparing clustering methods for market segmentation: A simulation study
Chad Vidden, Marco Vriens, Song Chen · Applied marketing analytics · 2016
This paper compares clustering methods on simulated data sets with different characteristics, such as degree of variance, whether there are overlaps between segments, the nature of the true clusters, and the absence/presence of categorical variables. Specifically, the paper compares K-means with latent class and ensemble analysis. The authors’ findings show that latent class analysis performs best in most cases, both in its ability to recover the true cluster members and in its ability to identify the correct number of clusters. Ensemble methods perform second best. K-means performs reasonably well with continuous variables. The current authors also tested the core member approach that can be applied on top of any clustering method, and found that it improved the identification of the correct cluster members.