Clustering
Clustering
What is clustering?
- Grouping similar data without labels
- Discovering hidden groups and patterns
- Used for customer segmentation
Adopting clustering changes who owns the question “How should we segment our customers?” Until now, the planning team first built buckets along axes it already knew, such as age group, region, or tier, and then pushed customers into them. Clustering reverses the order, so the data proposes the segmentation axes itself. Instead of “women in their 30s,” you get groups no one would have thought of in a meeting room, such as “late-night mobile shoppers who often add small purchases to their carts.” It is a tool that generates hypotheses rather than one that tests them, and that is the biggest difference practitioners notice.
Algorithms differ in character. K-means is fast, but a person must set the number of clusters K in advance, and it assumes round, blob-like groups. Hierarchical clustering lets you see how groups merge in a dendrogram, which suits exploring a few thousand records. DBSCAN uses density as its criterion and flags points that fit no cluster as outliers. Still, putting variables on the same scale during preprocessing affects results more than the choice of algorithm. If you feed in annual purchase amount (in the millions of won) and purchase count (single digits) as they are, the amount variable alone dominates the distance calculation. You can estimate a suitable K with the elbow method or silhouette score, but also consider that the number of groups your operations team can actually handle is usually 3 to 6.
The most common pitfall in practice is that clustering always gives you an answer. Even when the data has no clear structure, set K=5 and it will split the data into five groups anyway. The moment you attach names like 'value-seekers' or 'premium' to them, the whole team takes those names as fact and falls into confirmation bias. Before naming a cluster, verify the differences between clusters against external metrics such as repurchase rate and average order value. Also, K-means results shift with the initial values, so if you rerun it every month, customers who sat in last month’s 'cluster 2' may move to this month’s 'cluster 4', breaking campaign history and making performance comparison impossible. Once the segments are final, it is safer to freeze the criteria and assign only new customers, and to set a fixed retraining schedule, quarterly or twice a year.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Seungtae ParkAX instructor who teaches by building AI business development and marketing automation himselfA long-established technique in statistics and pattern recognition, and now the best-known method of unsupervised learning.
Consider a hypothetical case. Company A is a household-goods D2C company with 180 employees and annual revenue of around KRW 120 billion (≈ US$89,780,000). It was sending the same coupon email every week to all 420,000 members. Within six months, the email open rate halved from 18% to 9%. The CRM team lead kept saying, “Splitting by tier doesn’t help. Apart from VIPs, everyone’s the same.”
- Manager Kim of the data analytics team pulled nine variables from the last 12 months of purchase history, including visit frequency, average order value, category concentration, and coupon response rate, normalized the units, and ran 420,000 customers through K-means.
- When the elbow graph bent ambiguously between 4 and 7, the CRM team lead cut in: we only have three operators, so we can’t run seven. They settled on K=5, which had the second-highest silhouette score.
- One of the five clusters had an average order value 1.8 times the mean, yet a coupon response rate of just 3%, the lowest of all. Assistant Manager Lee of marketing said, 'We’ve basically been sending these people nothing but discount emails for six months,' and split them off as a new-product early-access group.
- The largest cluster, 140,000 customers, turned out to be a near-dormant segment more than 90 days past their last purchase, so coupons were replaced with reminder notifications timed to the repurchase cycle, and sending was cut from weekly to twice a month.
- Three months later, the email open rate rose from 9% to 21%, and coupon-linked sales grew 34% over the previous quarter. With send volume down 40%, messaging costs also fell from KRW 11 million (≈ US$8,230) to KRW 6.8 million (≈ US$5,090) per month.