Clustering: Find Groups Without Target Labels
A shop has many customer records but no predefined customer categories.
Clustering: Find Groups Without Target Labels explained visually
A shop has many customer records but no predefined customer categories. English captions are optional and start off. The complete transcript follows the written lesson.
What you will learn
A shop has many customer records but no predefined customer categories.
Grouping depends on a definition
A shop has many customer records but no predefined customer categories. Clustering can organize records according to a chosen notion of similarity. The groups are produced by the features, distance measure, and algorithm. They are not automatically natural categories that exist independently of those choices.
Understand the algorithm assumptions
K-means alternates between assigning points to centers and updating the centers. It favors compact groups under its distance-based objective and requires a chosen number of clusters. Scaling matters because large numerical ranges can dominate distances. Different initializations can produce different solutions. Other methods, such as hierarchical or density-based clustering, make different assumptions and may suit different structures.
Check whether groups are useful
Evaluate clusters for stability, separation, and usefulness to the project. Internal scores can help compare some choices, but they do not prove that a group has a meaningful real-world interpretation. Inspect representative and borderline examples. Avoid turning a cluster label into a personal or causal claim that the data do not support. When new data arrive, check whether the original grouping still describes them reasonably.
A worked example
If customers are represented only by total spending, a clustering method may mostly separate low and high spenders. Adding visit frequency can produce a different grouping. Neither result is automatically wrong, but each answers a different question about similarity.
Worked example
If customers are represented only by total spending, a clustering method may mostly separate low and high spenders. Adding visit frequency can produce a different grouping. Neither result is automatically wrong, but each answers a different question about similarity.
Quick check: test your understanding
Try it yourself: What would happen if a distance calculation combines spending in thousands of dollars with a feature ranging only from zero to one?
Check the reasoning
Spending could dominate the geometry unless scaling or a deliberately chosen distance accounts for the difference.
Does a cluster label prove that every member shares the same underlying motivation?
No. It only reflects similarity under the selected representation and algorithm.
How would you solve the practice case?
Spending could dominate the geometry unless scaling or a deliberately chosen distance accounts for the difference.
What does the worked example show?
If customers are represented only by total spending, a clustering method may mostly separate low and high spenders. Adding visit frequency can produce a different grouping. Neither result is automatically wrong, but each answers a different question about similarity.
Further reading
Machine Learning — Tom M. Mitchell.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 13 transcript206 words · Video transcript · Matches the published narration and captionsOpen transcript
A shop has many customer records but no predefined customer categories. Clustering can organize records according to a chosen notion of similarity. The groups are produced by the features, distance measure, and algorithm. They are not automatically natural categories that exist independently of those choices.
K-means alternates between assigning points to centers and updating the centers. It favors compact groups under its distance-based objective and requires a chosen number of clusters. Scaling matters because large numerical ranges can dominate distances. Different initializations can produce different solutions. Other methods, such as hierarchical or density-based clustering, make different assumptions and may suit different structures.
Evaluate clusters for stability, separation, and usefulness to the project. Internal scores can help compare some choices, but they do not prove that a group has a meaningful real-world interpretation. Inspect representative and borderline examples. Avoid turning a cluster label into a personal or causal claim that the data do not support. When new data arrive, check whether the original grouping still describes them reasonably.
If customers are represented only by total spending, a clustering method may mostly separate low and high spenders. Adding visit frequency can produce a different grouping. Neither result is automatically wrong, but each answers a different question about similarity.