Comparaison entre les indices de validité pour obtenir le cluster optimal dans le clustering.

Abstract

Clustering is a model that explores the similarities and intrinsic structures of data to group objects into clusters, without using pre-existing labels. There are many clustering methods, and one of the most commonly used is the k-means method. In this thesis, we propose a comparison between several validity indices by combining them with the k-means algorithm to determine the optimal number of clusters. The method involves varying the number of clusters within a predefined range [kmin, kmax] and extracting the best index values representing the final number of clusters. The compared indices are Calinski and Harabas index, Hartiga, Ball and Hall index, Dunn index, Silhouette index, Davies Bouldin index, Xie Beni index, WSJI, WB, and Bayesian Information Criterion. The experimentation and comparison of validity indices were performed on synthetic datasets. The results confirm the effectiveness of certain indices such as Calinski and Harabas indices, Silhouette index, Davies Bouldin index, and Sum of Squares (WB) among different datasets.