Can a Machine Discover Hidden Groups?
How an unsupervised algorithm like k-means turns feature vectors into discrete clusters, and why its verdict is real for some shapes and unreliable for others.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 8 scenes and explore, respond, and learn as you go.
Can a machine, given only raw data points with no labels, discover the hidden groups inside them?
Imagine 60 unlabelled points scattered on a page, with no legend or categories — could an algorithm sort them into meaningful groups on its own?
Humans need labels to recognise groups, yet clustering algorithms work blindly. The question is whether they find real structure or only an illusion of order.
A live 2D simulation lets learners scatter, wiggle, and arrange points while watching k-means partition emerge, colour-banding grow, and centroids migrate step by step.
A machine can reveal hidden groups, but only when those groups are shaped like roughly equal, roundish blobs in feature space — and the verdict always depends on the distance metric and k you choose.
Yes — if humans can eyeball clusters, surely an algorithm that measures distances between points can do it better and more consistently.
- Hierarchical clustering, DBSCAN, and spectral methods
- Cluster validation indices such as silhouette or gap statistic
- Dimensionality reduction techniques like PCA and t-SNE
- 01Sixty Points, Zero LabelsslideQuestion
Introduce the driving question: a scatter plot of 60 unlabelled points is shown, and learners are asked what structure, if any, a machine could find without being told the categories.
- The dataset has no labels, no legend, and no hint of how many groups exist
- Only the (x, y) coordinates are available to any algorithm
- The challenge is to discover hidden groups from geometry alone
- 02Your First GuessquizPrediction
Learners commit to whether a machine can recover hidden groups from the 60 unlabelled points before any algorithm runs.
- One independent choice: yes or no, with reasoning
- The guess is locked in before any evidence is shown
- 03Watch K-Means Partition the PointsinteractiveEvidence
A 2D scatter is animated step by step: random centroids are placed, every point is coloured by nearest centroid, centroids recentre, and the bands stabilise into k regions.
- Each iteration reassigns points to their nearest centroid
- Centroids migrate to the mean of their assigned points
- The final colour bands reveal the algorithm's partition
- 04Stress-Test the AlgorithminteractiveEvidence
Learners reshape the dataset — wide and narrow blobs, overlapping rings, unequal sizes — and re-run k-means to see where it succeeds, distorts, or invents fake groups.
- Tight, round blobs are recovered cleanly
- Concentric rings are sliced radially instead of being recognised
- Unequal clusters are split or merged against the real structure
- 05Why It Works — and Why It FailsslideExplanation
Explain that k-means minimises within-cluster sum of squared Euclidean distances, so it favours convex, similarly sized blobs; non-spherical or overlapping groups violate that implicit assumption.
- The objective function rewards compact, equally sized balls
- Distance metric and k are user choices, not discovered truths
- Different seeds can yield different partitions of the same data
- 06Apply It to a New DomaininteractiveTransfer
Learners change the simulation to a customer-feature scatter (age vs. spend) and decide whether k-means is the right tool, choosing k and justifying the partition.
- Translate the geometry intuition into a business context
- Defend a chosen k against alternatives
- Recognise when k-means imposes structure that may not exist
- 07Where the Answer Stops Being YesslideBoundary
Define the boundary: k-means discovers hidden groups only when groups are roughly globular, similarly sized, and separable in Euclidean space; beyond that, the 'discovery' is the algorithm's projection, not a fact.
- Boundaries include rings, manifolds, and nested density
- High dimensionality stretches Euclidean distance until it loses meaning
- Outside these bounds, the machine is inventing structure rather than finding it
- 08Yes — Within Strict LimitsslideResolution
Resolve the driving question: a machine can discover hidden groups when the geometry matches the algorithm's assumptions; otherwise it imposes order rather than uncovering it.
- Yes, for compact, similarly sized blobs separated by clear gaps
- No, for rings, nested density, or arbitrary k choices
- The algorithm reveals structure, but the structure it reveals is partly its own
Discussion threads for a Stage aren't available yet.
