Comparison and Evaluation of Clustering Models

Nowadays there are a variety of clustering models and evaluation metrics have been proposed, however, there is still a lack of unified gold standard for the evaluation of these models, especially when evaluating the model performances and robustness. This paper proposes an innovative evaluation strategy for clustering models, and conducts the testing on five clustering algorithms: K-means, Hierarchical clustering, Leiden, Spectral clustering and DBSCAN. To visualize the effects of the clustering, the UMAP dimensionality reduction method is implemented to show the distribution of the clusters clearly. Meanwhile, we need to combine with 10 evaluation indicators to further evaluate the model performances. Moreover, by using the data with classification labels, the accuracy of these models was calculated. In these ways, the clustering models were comprehensively evaluated. Next, to test the robustness of the clustering models, this paper adds data noise with isometric standard deviation from 0 to 0.15 on real data set. The results show that K-means and Spectral clustering perform best in accuracy and clustering quality, especially spectral clustering shows strong robustness on noisy data and simulated data, while DBSCAN and Leiden are relatively insufficient in dealing with complex data.