Clustering & Segmentation: Choose the Right Number of Clusters in Interviews
A quick-commerce team is deciding whether a city has three natural demand zones or seven. Pick too few clusters and every neighbourhood gets the same assortment; pick too many and operations becomes unmanageable.
That is the real tension in clustering: the “right” number of clusters is not the prettiest chart - it is the smallest useful segmentation that is statistically defensible and operationally actionable.
- Clustering groups similar observations without a target variable; choosing k means deciding how many groups to create.
- The best k balances four tests: compactness, separation, stability and business actionability.
- The elbow method looks for the point where adding more clusters gives sharply lower incremental benefit.
- The silhouette score checks whether points are closer to their own cluster than to the nearest other cluster; closer to 1 is better.
- Never choose k from one metric alone. Validate with multiple k values, profiles, cluster sizes and business feasibility.
- A useful segment must be measurable, reachable, distinct and actionable - otherwise it is just a statistical pattern.
- Interview-safe answer: “I would shortlist k statistically, test stability, profile clusters, then choose the k that improves a business decision.”
Big Picture
Choosing the number of clusters is a translation problem: turn messy behavioural or operational data into a few groups that a manager can actually act on. The analytics team may test k = 2 to 10, but the business finally needs names, priorities and decisions - “price-sensitive repeat buyers,” “premium convenience seekers,” or “high-density quick-delivery catchments.”
Core Explanation
Clustering is unsupervised learning: there is no “correct answer” column. The algorithm groups observations based on similarity across selected features - for example recency, frequency, monetary value, product category mix, delivery location, app engagement or complaint behaviour.
k is the number of clusters. In k-means, you must choose k before fitting the model. In hierarchical clustering, you choose the level at which to cut the dendrogram. In Gaussian mixture models, you compare candidate numbers of components. In DBSCAN, you do not choose k directly, but you still influence group count through density parameters.
The key idea: more clusters almost always fit the data better, but not always the business better. If every customer becomes their own cluster, compactness is perfect and usefulness is zero.
The Four Questions That Decide k
The Practical Five-Step Process
Metrics to Track When Choosing k
Use metrics as evidence, not as a judge. A strong candidate k usually performs well on several measures and produces clusters that the business can understand.
Worked Example - Silhouette Score in One Minute
Suppose customer A is in Cluster 1. Its average distance from other customers in Cluster 1 is 2.0, so a = 2.0. Its average distance from the nearest other cluster is 5.0, so b = 5.0.
Silhouette for customer A = (b - a) / max(a, b) = (5.0 - 2.0) / 5.0 = 0.60. That is a good sign: customer A is much closer to its own cluster than to the nearest competing cluster.
Now compare average silhouette across candidate k values in a toy run:
If k = 3 also gives interpretable profiles and stable cluster sizes, it is the best recommendation. If k = 4 enables a valuable action, such as a separate premium service tier, you may still choose 4 - but you must justify the trade-off.
How Different Algorithms Affect the Choice
An Indian eyewear or fashion retailer can cluster customers using purchase frequency, average order value, product category, discount sensitivity and store-versus-app behaviour. The strategic point is not whether the model produces five clusters; it is whether those clusters support different actions such as premium frames, value bundles, reminder campaigns or store-level assortment decisions.
Definitions
Clustering: An unsupervised learning method that groups observations so within-group similarity is high and between-group similarity is low.
Number of clusters, k: The chosen count of groups into which the algorithm partitions observations.
Kotler and Keller: “Market segmentation divides a market into well-defined slices.”
Silhouette score: A measure of how close an observation is to its own cluster compared with the nearest other cluster.
Zepto: Choosing Useful Clusters in Quick-Commerce Operations
Zepto’s quick-commerce model shows why the right number of clusters must serve an operating decision, not just a statistical chart.
Zepto operates in a category where locality-level differences are central. A neighbourhood with dense apartment complexes, frequent top-up grocery orders and tight delivery windows behaves very differently from a lower-density locality with larger baskets and longer travel times.
The segmentation problem is practical: how many operating clusters should a city be divided into for decisions such as catchment design, assortment, rider planning, replenishment and promotions?
The move is to evaluate candidate clusters using both data and execution reality. Demand density, order timing, SKU mix, road connectivity, distance to dark stores and service constraints can create statistically visible groups. But the chosen number of clusters must also be manageable for teams that run inventory, pricing, delivery and local marketing.
The lesson: if k = 8 gives slightly better statistical separation but operations can only run four distinct playbooks, k = 4 may be the smarter managerial answer. Strong segmentation is not maximum granularity; it is decision-ready granularity.
How AI Changes Choosing the Number of Clusters
AI changes clustering in three concrete ways.
- Embeddings create richer features: Instead of clustering only transaction numbers, teams can cluster text reviews, support tickets, product descriptions or browsing journeys using embeddings. This makes segments more behavioural, but also harder to interpret.
- AutoML accelerates candidate testing: Modern ML platforms can test algorithms, k ranges, scaling methods and validation metrics faster. The student trap is to accept the best automated score without checking business meaning.
- LLMs help profile clusters: Once clusters are formed, an LLM can summarise top variables, generate possible segment names and draft business actions - but the analyst must verify that the names reflect actual data.
Student workflow: Load your clustering output table into ChatGPT or Claude with columns such as cluster size, average recency, frequency, spend, category mix and silhouette score. Ask: “Name each cluster, identify the business action for each, and flag any cluster that looks statistically weak or commercially unclear.” Then manually check the answer against the data.
Interview Relevance
“You have clustered customers for a retail company. How would you decide the right number of clusters?”
Use this sentence in interviews: “I would not choose k from the elbow alone; I would triangulate metrics with stability and then select the smallest number of clusters that supports distinct business actions.”
Common Mistake
The biggest mistake is saying, “I will use the elbow method and choose the elbow,” then stopping there. It costs candidates because it sounds mechanical and ignores whether clusters are stable, interpretable and usable. Fix: say, “Elbow is a shortlist tool; the final k must also pass silhouette, stability, profile clarity and actionability.”
What to Revise Next
Once clustering feels clear, revise supervised model evaluation. The natural next step is learning how model metrics change when the business problem changes.