Computational Thinking Classes in Montpellier

Chapter 128 – K-Means Clustering Explained

1. Supervised vs Unsupervised

Supervised models train on features X and labels y. Unsupervised models see only X and must discover structure. Clustering is a common unsupervised task: choose a number of groups K and let the algorithm form them.

2. Where Clustering Is Used

• Customer segments for marketing
• Unusual transaction groups in fraud analysis
• Grouping similar articles or images
• Browsing-behavior clusters for ads

3. Quality Without Labels

Without ground-truth labels, quality is judged by distances:

• Within-cluster distance — minimize (points in a group stay tight)
• Between-cluster distance — maximize (group centers stay apart)

4. What K-Means Does

K-Means partitions data into K clusters around moving centers called centroids. It is transparent and widely used as explainable unsupervised learning.

5. Four-Step Loop

1. Initialize K centroids (random or smarter schemes like k-means++)
2. Assign each point to the nearest centroid (often Euclidean distance)
3. Update each centroid to the mean of its assigned points
4. Repeat assign + update until stopping rules fire

6. When to Stop

Criterion	Meaning
Max iterations	Hit a fixed iteration budget
Centroid movement	Shifts fall below a small threshold ε
Stable assignments	No points change cluster between rounds

7. Strengths and Care

K-Means is fast and scales well, but results depend on starting centroids. Practical fixes: use k-means++ initialization, or run several random seeds and keep the run with the smallest within-cluster distance.