Silhouette Score

When you cluster data, the hard part is not just forming groups—it is judging whether those groups actually make sense. Silhouette Score gives you a simple way to ask: is each point closer to its own cluster than to neighboring clusters?

For each data point, the silhouette value compares two distances:

• a: the average distance from that point to other points in its own cluster
• b: the average distance from that point to points in the nearest different cluster

The score is computed as (b - a) / max(a, b), which puts it between -1 and 1. A value near 1 means the point fits well inside its cluster and is far from others. A value near 0 means it sits on a boundary. A negative value means it may have been assigned to the wrong cluster. The overall Silhouette Score is the average across all points.

In unsupervised learning, you usually do not have true labels, so internal checks like this are essential. The silhouette score helps compare clustering results from methods like K-Means, agglomerative clustering, or spectral clustering, and it is widely used to choose the number of clusters k. In customer segmentation, for example, a higher score suggests the customer groups are more distinct and internally consistent. In document clustering or image grouping, it helps reveal whether the model found real structure or just forced the data into arbitrary partitions.

The metric works best when clusters are fairly compact and well separated. It can be misleading for irregular shapes, varying densities, or algorithms like DBSCAN that discover non-spherical clusters. It also depends on the chosen distance metric, such as Euclidean or cosine distance. In practice, people often compute it with scikit-learn’s silhouette_score and inspect both the average score and a silhouette plot, because the plot shows whether some clusters are strong while others are messy.