고객 세분화를 위한 클러스터링: 처음부터 끝까지 살펴보기

Customer segmentation groups buyers by behaviour so that marketing, product, and customer-success teams can tailor their actions to each group. Typical signals include recency (days since last purchase), frequency (number of purchases), and monetary value (total spend) — the RFM framework. Clustering discovers these segments from data without needing predefined categories.

We use the classic Online Retail dataset (UCI ML Repository). It contains ~500k transactions with invoice date, customer ID, quantity, and unit price. Our first task is to load the data, drop rows with missing customer IDs, filter out returns (negative quantity), and compute the RFM features for each customer.

Recency: days since the customer's last purchase (smaller = more recent = better). Frequency: number of unique invoices. Monetary: total revenue generated. We compute these relative to a snapshot date (one day after the last transaction in the dataset) so recency increases with inactivity.

RFM features are often highly right-skewed: a handful of VIP customers dominate the monetary axis. Before scaling, apply a log transform (np.log1p) to compress the long tail. Clip extreme outliers beyond the 99th percentile to prevent a single whale customer from distorting all centroids.

K-Means uses Euclidean distance, so features must be on the same scale. After log-transforming, apply StandardScaler to centre each feature at zero with unit variance. Always fit the scaler on training data only — here the full RFM table since there is no separate test set for unsupervised learning.

Run the elbow and silhouette diagnostics on the RFM dataset to select k. For a typical e-commerce dataset you might see the elbow around k=4 or k=5, which corresponds to intuitive segments: champions, loyal customers, at-risk customers, and churned customers.

After selecting k, fit the final K-Means model and add the cluster labels back to the RFM DataFrame. This makes it easy to compute segment profiles and build customer-facing reports. The fit_predict method fits and returns labels in one call.

Compute the mean of the original (untransformed) RFM values for each cluster. This gives interpretable business profiles: Champions have low recency, high frequency, high monetary; Churned have high recency, low frequency, low monetary. Naming segments based on their profiles makes reports actionable.

Plot Frequency vs Monetary with colour coding for each segment. Add recency as point size to encode the third dimension visually. This chart is the deliverable that a marketing team can use to identify which customers to target for reactivation campaigns vs upselling campaigns.

After deploying the model, new customers get assigned by passing their scaled RFM vector through the same scaler and then calling km.predict. Never refit the scaler on new data — use the scaler fitted on the training RFM table to avoid shifting the feature space. The centroid positions remain fixed after fitting.

Clustering is a starting point, not an end. After profiling segments, the team should design targeted actions: send re-engagement emails to At-Risk customers, offer loyalty rewards to Champions, present upsell offers to Loyal customers. Track conversion rates per segment to measure the ROI of segmentation. Periodically retrain the model as customer behaviour evolves over time.

In this lesson you learned: RFM (Recency, Frequency, Monetary) features are the standard building blocks for customer segmentation, log transformation and StandardScaler make skewed RFM features suitable for K-Means, and segment profiling translates cluster numbers into actionable business labels like Champions and At-Risk. Next up we explore PCA — a technique for reducing high-dimensional data to its most informative components.