Learn
A4.3.4 | CLUSTERING (HL ONLY)
01 | FIND GROUPS WITHOUT BEING GIVEN THE ANSWERS
An online shop has records of how often customers order and the typical value of their baskets. It does not have a reliable label saying which kind of customer each person is. Could the purchasing patterns suggest useful groups?
Clustering is an unsupervised learning approach that groups observations using similarities in their features. The algorithm is not trained to reproduce known customer-category targets. Instead, it finds structure under the representation and grouping rules we choose.
This HL lesson uses fictional purchasing data to show the process. We will work through K-means as one concrete technique, then consider alternative approaches and why discovered groups need careful interpretation.
02 | WHAT UNSUPERVISED MEANS HERE
In supervised classification, each training example has a known target label, such as Type A or Type B. The model learns or uses those examples to predict a label for a new input.
In clustering, the inputs are supplied without known target-group labels guiding the fit. The algorithm assigns group IDs such as Cluster 1 and Cluster 2 based on feature patterns. Those IDs are outputs of the grouping process, not previously verified customer types.
Unsupervised does not mean no human choices. People still define the purpose, select features, clean and scale them, choose a method and inspect whether the results are useful.
03 | CHOOSE FEATURES THAT DESCRIBE THE QUESTION
For customer segmentation, possible features include purchase frequency, average basket value and the proportions of spending on different product categories. Each feature contributes a coordinate describing one customer’s recorded behaviour.
A customer identifier is useful for joining the result back to the correct account, but a sequential account number does not normally provide a meaningful numerical distance between customers. Keep identity separate from clustering inputs.
The observation period matters too. Comparing one person’s full year of orders with another person’s first week can create misleading differences. Define consistent windows and investigate missing or incomplete histories before clustering.
04 | READ THE UNLABELLED PURCHASING DATA
| Customer | Frequency score x | Basket-value score y | Known segment label |
|---|---|---|---|
| C1 | 1 | 1 | Not supplied |
| C2 | 1 | 2 | Not supplied |
| C3 | 2 | 1 | Not supplied |
| C4 | 4 | 7 | Not supplied |
| C5 | 5 | 7 | Not supplied |
| C6 | 5 | 8 | Not supplied |
| C7 | 8 | 2 | Not supplied |
| C8 | 8 | 3 | Not supplied |
| C9 | 9 | 2 | Not supplied |
Both input descriptors are fictional comparable-scale scores from 0 to 10. Higher frequency scores represent more frequent purchasing; higher basket-value scores represent larger typical baskets. They are not literal numbers of orders or currency amounts.
The customer IDs identify rows only. The final column reminds us that no correct segment label is provided to the clustering procedure. A plot may suggest groups, but those visual impressions still depend on the two selected inputs.
05 | SIMILARITY DEPENDS ON THE REPRESENTATION
Two observations are similar if they are close under the chosen features and comparison rule. In a two-number representation, a common distance is Euclidean distance: square each coordinate difference, add the squares and take the square root.
C1 = (1,1) and C2 = (1,2) are one unit apart. C1 and C7 = (8,2) have distance √(7² + 1²) = √50 ≈ 7.071. Under these descriptors, C2 is much closer to C1.
This does not prove the customers are similar in every respect. They might buy different products, have different histories or behave differently outside the observation window. Similarity is always defined with respect to the supplied representation.
06 | SCALE BEFORE COMPARING UNLIKE MEASUREMENTS
Suppose raw basket value is recorded in hundreds of currency units but purchase frequency ranges only from 0 to 10. Numerical basket differences could dominate an unscaled distance simply because of magnitude.
Use appropriate training or reference-data preprocessing, such as standardization, and apply the fitted transformation consistently. The interactive dataset already uses comparable-scale descriptors to keep the calculation visible.
Scaling is itself a modelling choice. It does not make irrelevant features informative, remove every effect of an outlier or guarantee meaningful segments. If the goal is to group future customers, learn preprocessing from the fitting data rather than refitting it separately for every new batch.
07 | K-MEANS STARTS WITH A REQUESTED NUMBER OF GROUPS
K-means represents each cluster by a centroid, its mean coordinate position. The configured number k states how many clusters the procedure is asked to form.
For k = 3, begin with three initial centroid positions. Assign observations to their nearest centroid, recompute each centroid from its assigned members and repeat. The procedure aims to reduce within-cluster squared distances.
K in K-means is not k in K-nearest neighbours. K-means forms groups without target labels; K-NN consults k labelled nearby examples to predict a class.
08 | ASSIGN EACH OBSERVATION TO A CENTROID
For one worked assignment, use centroids A = (1,1), B = (2,1) and C = (4,8). These intentionally uneven starting positions are selected to make subsequent movement visible.
Customer C3 = (2,1) is distance 1 from A, distance 0 from B and distance √53 ≈ 7.280 from C. It initially belongs to B because B is nearest.
| Customer C3 to centroid | Squared distance | Euclidean distance |
|---|---|---|
| A = (1,1) | (2−1)² + (1−1)² = 1 | 1 |
| B = (2,1) | 0² + 0² = 0 | 0 |
| C = (4,8) | (2−4)² + (1−8)² = 53 | √53 ≈ 7.280 |
Each point is assigned using input distances, not a majority vote of known labels. If two centroids are exactly equally near, a defined tie rule is needed; this demonstration selects the lower-numbered cluster.
09 | RECOMPUTE A CENTROID BY AVERAGING ITS MEMBERS
After the first assignment with those starting positions, Cluster B contains C3 = (2,1), C7 = (8,2), C8 = (8,3) and C9 = (9,2).
Average the x coordinates: (2 + 8 + 8 + 9) / 4 = 6.75. Average the y coordinates: (1 + 2 + 3 + 2) / 4 = 2. The updated centroid is therefore (6.75, 2).
A centroid need not coincide with an actual customer. It summarises coordinates, not a real person whose behaviour must exist. Replacing an observation with a centroid would change the original dataset; K-means instead retains observations and updates the group representation.
10 | REASSIGN AFTER THE CENTROIDS MOVE
The first update moves Cluster A’s centre to (1,1.5), B’s to (6.75,2) and C’s to approximately (4.667,7.333). Now reconsider C3 = (2,1).
Its distance to A becomes √(1² + 0.5²) ≈ 1.118. Its distance to B becomes √(4.75² + 1²) ≈ 4.854. It changes from B to A.
That is why K-means repeats the process. The closest centre can change after the previous membership assignment has moved the centres. A single assignment from initial guesses is not the completed grouping.
11 | REPEAT UNTIL A STOPPING RULE IS MET
Alternate assignment and mean updates until the group memberships or centroid positions stop changing sufficiently, or a configured iteration limit is reached. A complete cycle contains an assignment stage and an update stage.
In the three-cluster worked example, the stable memberships are C1–C3 together, C4–C6 together and C7–C9 together. Their centroids are approximately (1.333,1.333), (4.667,7.333) and (8.333,2.333).
Convergence means the procedure has reached its stopping condition. It does not prove the globally best grouping or the existence of exactly three natural customer types. Different starting positions can produce different local solutions.
12 | STEP THROUGH ASSIGNMENT AND CENTROID MOVEMENT
Choose two or three clusters and a starting layout, then reveal one stage at a time. The plot and table update from real calculations on the nine fictional observations. Cluster IDs are arbitrary; compare memberships rather than assuming an ID has a fixed meaning.
Changing a selector starts a fresh run with that setting. The lower button reveals the next stage without making you scroll back up.
For k = 3 with uneven centres, watch C3 switch clusters after the first centroid update. During an update stage, the table keeps the previous memberships until the next assignment stage. The displayed squared-error sum always refers to the memberships and centres currently shown.
13 | UNDERSTAND WHAT THE OBJECTIVE MEASURES
K-means aims to minimise the within-cluster sum of squared distances from observations to their assigned centroids. It is often called inertia. Smaller values indicate tighter groups under that representation, for a specified comparison.
Assignment chooses the nearest centre; updating to a mean minimises squared distances within each nonempty fixed group. These stages do not increase that objective under the ordinary procedure.
This value is not classification accuracy and has no verified target labels in its calculation. A low objective does not prove that a group is useful, fair or causally meaningful. Comparing values across different scaling choices is especially difficult because the distance units have changed.
14 | WHY NOT INCREASE K UNTIL EVERYONE HAS A GROUP?
More clusters generally permit a lower best within-cluster squared-distance sum. With one centroid for each of nine distinct observations, the training sum can reach zero. That does not create nine useful shared customer segments.
A small k may combine meaningfully different behaviours. A very large k can fragment the sample into groups too small to be stable or useful. The goal is a useful structure, not simply the lowest raw training sum.
Compare candidate groupings using domain relevance, stability, separation and the intended use. An elbow plot can suggest where improvements flatten, but it need not have one clear elbow. Scores such as silhouette compare cohesion and separation; neither supplies a universal correct k.
15 | STARTING CENTRES CAN CHANGE THE RESULT
K-means can settle into a local solution, so its initial centres matter. Run it from multiple suitable starting positions and compare the resulting objective and stability. More deliberate initialisation methods can help, but do not guarantee a globally optimal answer.
The two teaching layouts let you compare starting paths. They may reach the same final partition with different cluster numbers or different numbers of stages. They are not promised to produce different final partitions every time.
A fixed random seed can make an implementation’s choices reproducible. Reproducibility means you can repeat the run; it does not establish that the chosen representation or result is correct.
16 | WHAT HAPPENS IF A CLUSTER HAS NO MEMBERS?
During some runs, a centre may receive no observations. Its new mean cannot be computed from an empty set. An implementation therefore needs an empty-cluster policy, such as reinitialising that centre.
The teaching activity retains an empty centre at its previous position and reports this if it occurs. Its supplied starting layouts are chosen for clear nonempty examples. This policy is a demonstration choice, not a universal requirement for K-means.
Distance ties also need a reproducible rule. Both cases show why a description should include the algorithm’s main process while recognising that practical implementations need explicit handling of edge cases.
17 | TURN GROUPS INTO CAUTIOUS BUSINESS DESCRIPTIONS
| Observed three-group pattern | Possible descriptive summary | What it does not establish |
|---|---|---|
| Lower frequency, lower basket score | Less frequent smaller-basket purchasers in this window | Low customer value for all time |
| Middle frequency, higher basket score | Customers with larger typical baskets in this window | A known demographic or cause |
| Higher frequency, lower basket score | More frequent smaller-basket purchasers in this window | A permanent identity or guaranteed preference |
The algorithm provides memberships and feature summaries. A human can attach descriptive names afterwards, but those names are interpretations, not discovered ground truth.
A shop might investigate whether different communications or stock planning are helpful for the groups. Any intervention needs appropriate evaluation; a cluster alone does not prove an offer will cause a desired outcome. Avoid turning purchasing behaviour into unsupported assumptions about people.
18 | K-MEANS IS ONE CLUSTERING TECHNIQUE
| Approach | Main idea | Useful distinction |
|---|---|---|
| K-means | Alternate nearest-centroid assignment and coordinate means | Requires a requested k and works naturally with numerical mean/distance structure |
| Hierarchical clustering | Build nested groups by successively merging or splitting observations | A dendrogram can show group structure at different levels |
| Density-based clustering | Find connected dense regions under a neighbourhood rule | Can identify irregular groups and leave some observations as noise |
Different methods encode different ideas of a group. Standard K-means favours compact groups under Euclidean geometry and assigns every observation to a centroid, including unusual ones. Density-based approaches may leave noise unassigned.
These methods do not all use centroid means or require the same settings. K-means is the worked example here, not the definition of all clustering. No one method is best for every shape, density or feature type.
19 | OUTLIERS AND SHAPES CAN MISLEAD
A very large basket-score outlier can pull a mean centroid away from the main observations. Investigate whether it is a genuine high-spend customer, a unit error or a different population before deciding how to handle it.
Two curved or interlocking groups may not be represented well by nearest-centroid partitions. Groups with very different sizes or spreads can also create difficulties for standard K-means.
A clustering algorithm can still return a partition when the data have no clear natural grouping. Output alone is not evidence of a hidden truth. Inspect the original features, the grouping assumptions and whether a different representation or method is justified.
20 | EVALUATE CLUSTERS WITHOUT INVENTING CORRECT LABELS
Without known target-group labels, ordinary classification accuracy, precision and recall cannot simply be calculated as if a correct segment existed for every customer. Assess the structure using suitable internal measures and domain checks.
Check whether members are reasonably similar under meaningful features, whether groups are separated, whether results remain useful across resampling or time windows and whether small representation changes radically alter them.
If independent reference labels exist for a particular evaluation, external comparisons are possible. Compare memberships with measures that respect arbitrary cluster IDs; Cluster 1 and Cluster 2 can swap names without changing the partition. For a deployment task, also evaluate the downstream benefit rather than only an attractive plot.
21 | ASSIGNING A NEW CUSTOMER IS NOT THE SAME AS REFITTING
With a fitted K-means model and fixed preprocessing, a new customer can be assigned to the nearest existing centroid. That uses the learned group structure without recomputing every centre.
Refitting instead recomputes groups from a dataset, possibly changing centres and memberships. Simply assigning a new customer does not automatically update the stored centroids.
A customer far from every centre still receives a nearest-centroid assignment under the basic rule. Monitor distances and changed purchasing conditions; the assigned group does not prove the new case is well represented. Other clustering methods may handle new observations differently.
22 | COMPARE CLUSTERING, CLASSIFICATION AND ASSOCIATION
| Task | What is supplied or sought | Purchasing example |
|---|---|---|
| Clustering | Group observations by feature similarity without target-group labels | Discover purchasing-pattern segments |
| Classification | Use known example labels to predict a predefined category | Predict whether an order will be returned |
| Association-rule learning | Find relationships between attributes or items occurring together | Explore which products appear together in baskets |
These can be used in related systems, but they answer different questions. A cluster is a group of observations; an association rule describes a co-occurrence relationship; a supervised class has a target meaning defined before prediction.
The next lesson develops association rules. For now, keep the central clustering chain clear: meaningful features → similarity comparison → group formation → careful interpretation and evaluation.
23 | CHECK THE GROUPING IDEA
Choose the main concept illustrated by each scenario. Focus on what the system actually does rather than the appearance of its chart.
24 | DESCRIBE THE PROCESS AS A CONNECTED STORY
Start with the task: “A shop represents customers using comparable purchasing features without known segment labels.” Then describe the grouping: “K-means assigns customers to nearby centroids and recalculates the centroids from their members, repeating until a stopping rule is met.”
Make the relationship clear: “Customers with similar feature coordinates tend to be grouped together under the distance rule. The resulting profiles can support investigation of different purchasing patterns.”
Finish with a limit: “Feature choice, scaling, k and initialisation influence the partition. Group IDs and names are not verified customer identities, so stability and usefulness must be checked.” This addresses the syllabus’s customer-base segmentation application without claiming that every dataset contains one correct segmentation.
Further reading: K-means stages and starting positions and clustering approaches and evaluation.