How Does Normalization Impact Clustering?

Normalization is becoming increasingly important with the surge in data complexity and volume in AI. Despite its significance, normalization often does not receive adequate attention. In this study, we evaluate various normalization methods on benchmark datasets across different clustering algorithms to understand their impacts on clustering. Contrary to the general assumption that density-based clustering methods are more robust to the effects of normalization than partitioning clustering methods, our results show that partitioning methods demonstrate greater efficiency and robustness under different normalization methods compared to density-based methods. Density-based methods may provide less reliable results, even when parameters are fine-tuned for well-structured data. Furthermore, density-based methods can produce significantly skewed clustering results for imbalanced data, regardless of the normalization methods used. However, partitioning methods can still achieve solid performance in handling imbalanced data. Our study sheds light on the development of more customized normalization methods for different clustering scenarios.