Frame the decision before choosing clusters
Clustering groups examples based on a chosen representation and similarity measure without requiring a target label for each example. It can help explore customer behavior, but cluster IDs are not automatically meaningful customer types or causal explanations. Define the intended action, the population, and the point in time for feature availability. Avoid including sensitive or post-decision attributes without a justified, governed purpose.
A retailer wants to tailor onboarding for new customers. List features available before onboarding and explain why “purchased after receiving the offer” cannot be an input to the initial segmentation.
Understand what k-means optimizes
K-means assigns examples to centroids using the configured distance and updates centroids to reduce within-cluster distance. The choice of k, initialization, distance, feature encoding, and scale affects the result. An outlier can pull a centroid; clusters with very different shapes or densities can be poorly represented. A feature measured in thousands can dominate a feature measured in single units under Euclidean distance unless scaling is considered.
Customers have annual spend from 0 to 10,000 and visits from 0 to 20. Explain why raw Euclidean distance may be driven mostly by spend and how a training-derived scaling method changes the interpretation.
Evaluate structure and stability
Inspect cluster sizes, centroids, within-cluster variation, and examples near boundaries. Compare several values of k and initializations, and assess whether assignments remain useful on later data. BigQuery ML can report the Davies–Bouldin index and mean squared distance for k-means. These internal measures describe aspects of cluster structure; neither establishes that a segment is useful, fair, or stable in production. Check the documentation for the model’s evaluation behavior, including which data is used.
One clustering solution has compact groups but one cluster contains 95% of customers. List two questions to ask about its usefulness and one test of stability across time or initialization.
Validate the downstream action
Translate each proposed segment into a concrete action and assess whether the action helps the intended population without unacceptable harms. A randomized comparison may be needed to measure an intervention’s effect. Monitor source feature quality, cluster sizes, assignment changes, and downstream outcomes; refresh or retrain under an explicit policy. Document that cluster labels can change between versions and should not be treated as permanent identities.
A team proposes a different welcome message for each segment. Describe a safe comparison, the primary outcome, a guardrail, and how to handle customers whose segment changes after a model update.
A k-means solution has a favorable internal clustering metric. What does that establish?
Keep exploring
- BigQuery ML k-means model options ↗
- BigQuery model evaluation overview ↗
- Evaluate clustering results ↗
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.