High-Level Synthetic Data Generation with Data Set Archetypes

Cluster analysis relies on effective benchmarks for evaluating and comparing different algorithms. Simulation studies on synthetic data are popular because important features of the data sets, such as the overlap between clusters, or the variation in cluster shapes, can be effectively varied. Unfortunately, curating evaluation scenarios is often laborious, as practitioners must find lower-level geometric parameters (like cluster covariance matrices) to match a higher-level scenario description like "clusters with very different shapes." To make benchmarks more convenient and informative, we propose synthetic data generation based on data set archetypes. In this paradigm, the user describes an evaluation scenario in a high-level manner, and the software automatically generates data sets with the desired characteristics. Combining such data set archetypes with large language models (LLMs), it is possible to set up benchmarks purely from verbal descriptions of the evaluation scenarios. We provide an open-source Python package, repliclust, that implements this workflow. A demo of data generation from verbal inputs is available at https://demo.repliclust.org.