Set 10 — k-means and customer segmentation
After lecture 11 · 3 November · about 40 minutes
Data: customers.csv — 6 000 customers, RFM-style. Columns: recency, frequency, monetary, tenure, returns, channel_online.
Before you start: call set.seed(10) right before each kmeans() unless a question names another seed — the Blackboard test checks your numbers against it.
1 · Scale, and say why
Run kmeans(customers[, 2:7], centers = 4) on the raw columns, then on scale()d columns. Compare the cluster sizes.
Which variable dominates the unscaled solution, and what is it about that variable that lets it dominate? Answer in one sentence that names the units.
2 · Choose k without the elbow
Plot total within-cluster sum of squares for k = 1:10, with kmeans(scale(customers[, 2:7]), k, nstart = 25). Then compute the average silhouette for k = 2:10 (cluster::silhouette, on the same scaled distances).
The two curves disagree, or one of them is ambiguous. Say which k you would take to a marketing director and what you would say when they ask why. An elbow you cannot see is not an argument.
3 · Stability is evidence
Run k-means with k = 4 on the scaled columns twice, with nstart = 1 and seeds 1 and 2. How many customers change cluster? Compare tot.withinss. Now with nstart = 50.
What does nstart actually do, and why does the answer to this exercise justify never leaving it at 1?
4 · Name the segments In class
We discuss this one in lecture 13, on 13 Nov — bring your answer.
Take the k = 4 solution. Compute the mean of each variable per cluster.
Give each cluster a name a marketing team could use, and one action per segment. If two clusters resist being told apart, say so — that is a finding, not a failure.
5 · The uncomfortable question In class
We discuss this one in lecture 13, on 13 Nov — bring your answer.
A colleague runs the same script with k = 7 and gets seven “meaningful” segments. They are not lying: the clusters exist and the means differ.
Write three sentences explaining why more clusters always look meaningful, and what evidence would distinguish a real segment from an arithmetic one.
Check yourself on Blackboard
Open the test Set 10 on Blackboard, enter your numbers, and read the feedback as soon as you submit. Not graded; as many attempts as you like.
This set checks: scaling, the difference between an objective that always improves and a statistic that does not, and the habit of defending a choice.