Evaluation datasets for clustering performance under compact and over-intermingled distributions

Abstract

This thesis considers two classes of synthetic datasets, namely compact-isolated and overlapping-intermingled clusters. The datasets are generated using fixed random seeds and predefined distribution parameters to ensure reproducibility and consistency. Each dataset represents a smart grid scenario in which smart meters are randomly distributed within an environment, and each meter must be assigned to an appropriate base station using the proposed hybrid metaheuristic scheme. The compact-isolated datasets, characterised by high intra-cluster similarity, provide a controlled and simplified setting for baseline benchmarking and clear interpretation of clustering performance. In contrast, the overlapping-intermingled datasets exhibit low cluster separability, creating a more challenging environment with greater noise, overlap, and variation in cluster density, shape, and size. These datasets therefore better reflect real-world conditions and are used to evaluate the robustness of the algorithms. For both dataset types, experiments are conducted in a two-dimensional space at three complexity levels: 4, 9, and 16 clusters, with corresponding dataset sizes of 400, 900, and 1600 data points.