Lesson 5 of 10

Choose Useful Clusters

K-means can return different results on the same data when initial centers change. A single run is not necessarily a unique correct grouping.

Choose initialization and group count separately

• Multiple starts: Keep K fixed, try different centers, and compare squared-distance sums. This can avoid poor starts without guaranteeing the global optimum.
• K-means++: Points farther from existing centers are more likely to become the next center. This helps spread out initialization.
• Vary K: Compare improvements in the objective alongside the practical purpose. Increasing K tends to reduce the objective, so its minimum alone is insufficient.

For example, 20 tidy customer groups may be impractical if a support team can offer only three distinct service approaches.

Compact within, separated between

A silhouette score compares a point’s distance to its own group with its distance to another group. Scores generally range from -1 to 1; higher values indicate better separation under the chosen distance. Evaluation normally requires at least two groups and fewer groups than data points.

A score favoring round groups may misjudge curved ones. Visualize the points and inspect what members have in common as well. The scikit-learn clustering guide compares assumptions and limitations.

Check your understanding

K=10 gives a smaller squared-distance sum than K=3. Does that prove K=10 is better?