10.7 Clustering and segmentation

Clustering is unsupervised. It does not predict a known target. It groups observations according to features and a similarity rule selected by the analyst.

A defensible clustering analysis requires:

• a clear reason for segmentation;
• an appropriate unit, such as one row per member rather than one row per transaction;
• selected features that represent meaningful behaviour or value;
• scaling when variables use different units;
• comparison of several cluster solutions;
• assessment of stability and separation;
• useful descriptions based on observed features; and
• a practical action that differs across segments.

K-means seeks groups that minimize within-cluster squared distance from cluster centres. It works best for roughly compact clusters in appropriately scaled numeric features. Hierarchical clustering can reveal nested structure. Other methods may be preferable for irregular shapes, mixed variable types, or noise.

Worked Example: Interpreting NVRW Member Segments

NVRW constructs one row per member with first-year visit frequency, program diversity, average contribution margin, digital-service use, location diversity, and twelve-month retention eligibility. After scaling the numeric features, the team compares several k-means solutions and examines stability across resamples.

One useful solution identifies frequent multi-service members, program-focused households, occasional drop-in users, and recent low-engagement members. The labels describe observed behaviour. The team does not infer income, motivation, or preferences that are absent from the data.