Fitness Function Comparison for Unsupervised Feature Selection with Permutational-Based Differential Evolution

This paper presents a comparative study of the performance of an unsupervised feature selection method using three evaluation metrics. In the existing literature, various metrics are used to guide the search for a better feature subset and evaluate the resulting data clusterization. Still, there is no well-established path for the unsupervised wrapper-based approach as for the supervised case. This work compares three metrics to guide the search in a permutational-based differential evolution algorithm to feature selection: the Silhouette Coefficient, the Kalinski-Harabasz Index, and the Davies-Bouldin Score. The experimental results indicate that no metric performed better when applying the feature selection process to thirteen datasets. Nevertheless, a clear tendency to select small subsets is observed. Furthermore, in some cases, performing the feature selection decreased the performance compared to the complete dataset.