Hands-On Unsupervised Learning with Python

Calinski-Harabasz score

Let's suppose that a clustering algorithm has been applied to a dataset X containing M samples in order to segment it into n_c clusters C_i represented by a centroid μ_i ∀ i = 1..n_c. We can define the Within-Cluster Dispersion (WCD) as follows:

If x_i is an N-dimensional column vector, X_k ∈ ℜ^{N × N}. It's not difficult to understand that WCD(k) encodes the global information about the pseudo-variance of the clusters. If the maximum cohesion condition is met, we expect a limited dispersion around the centroids. On the other hand, WCD(k) can be negatively influenced even by a single cluster containing outliers. Hence, our goal ...