A cluster without a name is not a result

🏢 The situation

Max is preparing next year's strategy:

"Marketing splits customers into 'new' and 'old' — that's the stone age. I want to know what customer types we really have, to build different products and communications. There are no labels — nobody knows the 'correct' segments. Is this possible?"

Lena:

"This is unsupervised learning. There is no target and no 'accuracy': segmentation quality is measured by whether the business can use it. Build RFM features, cluster, and above all — describe the segments in words. Clusters without interpretation are garbage."

🎯 Your task

1. Understand K-Means: how it works, how to pick the number of clusters.
2. Prepare features for clustering (scaling is critical!).
3. Interpret the segments and hand them to the business.

---

📚 Theory

K-Means in 4 steps

1. Pick K random centers.
2. Assign every point to the nearest center.
3. Recompute the centers as the means of their points.
4. Repeat 2–3 until convergence.

Properties: fast, simple; finds convex clusters of roughly equal size; sensitive to feature scale and outliers; the result depends on initialization (cured by n_init=10, k-means++).

How to choose K

• Elbow method: plot inertia (sum of squared distances to centers) vs K; look for the "bend".
• Silhouette score (−1…1): how much closer a point is to its own cluster than to the nearest other; higher is better.
• The business criterion (the main one): how many segments can the team actually serve? 4–7 is usually the ceiling for marketing.

Scaling is not optional

K-Means measures Euclidean distance. If "revenue" is in thousands and "purchase frequency" in single digits, clustering collapses to revenue. StandardScaler is mandatory; log-transform skewed money amounts.

RFM — the classic of customer analytics

• R (Recency) — days since the last purchase;
• F (Frequency) — number of purchases over the period;
• M (Monetary) — total spend.

Plus behavioral: share of discounted purchases, category diversity, average basket.

Alternatives to K-Means

• DBSCAN — arbitrary-shape clusters, finds outliers itself; needs no K but is sensitive to density parameters.
• Hierarchical — a dendrogram; you see the nesting of segments.
• Gaussian Mixture — "soft" probabilistic clusters.

---