How a Digital Street Decides Who's Neighbors
That a digital street is a chosen distance metric, and that the clusters you find are the shapes that metric carves out — not an objective truth about the data.
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.
What does 'clustering on a digital street' actually mean, and why does the same data form different neighborhoods depending on how the street is built?
Walk down any real street and you see houses clustered by size, era, and price. Now picture a street that exists only as numbers — what clusters there?
We assume clustering is a human act of grouping by eye. On a digital street, no eye exists — so what does 'near' even mean when the street is made of coordinates?
Watch identical points land in different neighborhoods depending on which 'street' metric you walk them down, and see the silhouette of each cluster tighten or blur in real time.
A digital street is a distance metric, and clustering is whatever shape that metric happens to carve out.
If the data is the same, the clusters should be the same — clustering reveals a real grouping, independent of the algorithm.
- Hierarchical clustering trees in detail
- DBSCAN density parameters
- Spectral and Gaussian mixture variants
- Real-world deployment pipelines
- 01A Street With No BricksslideQuestion
Open with the analogy: a real street has houses, lampposts, and curbs. A digital street has only coordinate pairs and a ruler. Frame the driving question visually with two simple scatter plots side by side.
- Real street: visible neighbors
- Digital street: coordinates + a distance function
- Driving question: what counts as 'near'?
- 02Commit Before You ComputequizPrediction
Ask the learner to predict which of three clustering outcomes is the 'true' one before any algorithm runs.
- One independent prediction
- 03Walk the Same Data Down Two StreetsinteractiveEvidence
Simulation that lets the learner run k-means on identical 2D points while toggling between Euclidean distance (straight-line ruler) and Manhattan distance (taxicab grid). Show how the cluster boundaries redraw themselves.
- Identical data, different distance
- Watch centroids snap to new homes
- Boundaries shift visibly
- 04Watch the Silhouette Tell on ItselfinteractiveEvidence
Visualization that recomputes a silhouette score as the learner drags k from 2 to 6 on the same dataset, revealing when the metric prefers compact vs stretched groups.
- Silhouette rises and falls with k
- Compact groups score high
- Stretched groups score low
- 05Why the Street Builds the NeighborhoodslideExplanation
Explain that k-means assigns a point to whichever centroid is shortest under the chosen metric. Euclidean geometry draws straight-line Voronoi cells; Manhattan geometry draws diamond cells. The data is identical; only the ruler changed.
- Assignment rule uses the metric
- Voronoi cells follow the ruler
- Same data, different neighborhood shapes
- 06When the Street Lies About the DataslideBoundary
Show a ring-shaped dataset where k-means with Euclidean distance fails — the metric assumes globular neighborhoods, so a concentric ring gets chopped into wedges. The street is wrong for this terrain.
- Euclidean assumes globular clusters
- Ring data breaks the assumption
- Wrong metric = wrong neighborhood
- 07Cluster a New Shape on the Old StreetinteractiveTransfer
Give the learner a fresh non-globular dataset (two interleaving moons) and ask them to predict whether changing the metric alone rescues the clustering — or whether a new street is needed.
- Apply to a changed shape
- Metric change is not always enough
- Decide what to test next
- 08The Street Is the DecisionslideResolution
Resolve the driving question: a digital street is a distance function, clustering is the partition that distance function makes compact, and 'true clusters' are never found — only chosen, by the ruler you pick.
- Street = metric
- Cluster = what's compact under that metric
- Truth is chosen, not found
Discussion threads for a Stage aren't available yet.
This path ends here.
