🤖 AI / ML

Clustering

clustering · cluster analysis · k-means clustering · customer clustering · unsupervised segmentation · clustering algorithm

Definition

Clustering is the task of splitting a set of objects into non-overlapping groups (clusters) that maximise similarity inside each group and difference between groups.

How clustering works

Clustering is an unsupervised learning task: the algorithm finds structure in the data on its own,
with no rules set in advance. The single criterion is that objects inside a cluster should resemble
each other, while objects from different clusters should differ.

The process, using k-means as the example:

K, the number of clusters, is chosen
K centroids are picked at random
Every object is assigned to its nearest centroid
Centroids are recomputed as the mean point of their cluster
Steps 3 and 4 repeat until convergence

Clustering algorithms

Algorithm	How it works	When to use it
K-means	Splits by distance to centroids	Customer segmentation by RFM, AOV
DBSCAN	Groups dense regions	Anomaly detection, outlier segments
Hierarchical	Builds a tree of nested clusters	Assortment taxonomy, category analysis
Gaussian Mixture	Probabilistic cluster membership	Soft segmentation: a customer can belong to several clusters

Applications in e-commerce

Customer clustering surfaces behavioural segments without hand-written rules:

Cluster 1: high LTV, buys rarely, large basket   — "considered buyers"
Cluster 2: frequent purchases, low AOV           — "browsers"
Cluster 3: bought once, disappeared              — "one-off buyers"
Cluster 4: only buys on promotion                — "discount hunters"

Each cluster needs its own personalization strategy.

Product clustering addresses cold start: a new SKU with no view history is assigned to the
nearest product cluster by its attributes and starts receiving traffic from recommendations.

Tip: do not read clusters mechanically. K-means produces mathematical groups; their business
meaning is decoded by analysts. Always validate a cluster by asking what these 50,000 customers
actually have in common.

Limitations

K-means is sensitive to feature scale — normalise the data before clustering
Clusters are unstable under random initialisation — use k-means++ or several runs
The optimal K is rarely obvious; the elbow method gives a hint, not an answer
Clusters are a snapshot in time; behaviour shifts, so clusters have to be refreshed

Frequently asked questions

How is clustering different from segmentation?

Segmentation is the broader term. Segments can be defined by business rules, for example VIP customers = LTV above 100K. Clustering is an algorithmic method where groups emerge from the data automatically, with no criteria specified in advance. Clustering is often the foundation for the segmentation that follows.

What is k-means and how do you choose K?

K-means splits data into K clusters by minimising the average distance from each point to its cluster centroid. K is usually picked with the elbow method: plot inertia against K and take the point where the curve bends. In practical e-commerce segmentation, K is typically between 4 and 12.

How is product clustering used in recommendations?

Product clusters help with the cold start problem. A new item can be assigned to the nearest cluster by its attributes and start being recommended to the audience that has historically bought from that cluster — without any accumulated interaction history for the specific SKU.