Why can a harmless-looking unit change alter a model?
Start here
Why can a harmless-looking unit change alter a model?
Distance- and gradient-based methods react to numerical scale. Changing kilometres to metres should not make one feature semantically more important, yet raw magnitudes can do exactly that.
Building interactive view…
Understand
Build the mental model
Distance- and gradient-based methods react to numerical scale. Changing kilometres to metres should not make one feature semantically more important, yet raw magnitudes can do exactly that. Fit scaling only on training data and include it inside a pipeline. Inspect transformed distributions and remember that scaling does not fix outliers or skew by itself.
Click a stage to inspect what happens, what changes, and what should be checked before moving on.
Stage 1
Raw feature scales
Identify exactly what enters this stage: its shape, type, scale, units and any missing or invalid values that could alter the next operation. Technical context for Scaling & Distance: Standardisation centres and rescales by training mean/std; Min–Max scaling maps a range; robust scaling uses resistant statistics. Tree splits generally do not require scaling.
Practitioner checkpoint: Fit scaling only on training data and include it inside a pipeline. Inspect transformed distributions and remember that scaling does not fix outliers or skew by itself.
What happens if…?
Break the assumption deliberately
Multiply one feature by 1,000 and watch KNN/SVM geometry change before and after scaling.
Move the control and explain what you expect before reading the visual.
Technical lens
Formalise what the visual is doing
Standardisation centres and rescales by training mean/std; Min–Max scaling maps a range; robust scaling uses resistant statistics. Tree splits generally do not require scaling.
Technical questionUse a tiny case to make the mechanism observable. Standardisation centres and rescales by training mean/std; Min–Max scaling maps a range; robust scaling uses resistant statistics. Tree splits generally do not require scaling. Verify one intermediate quantity, state change or mapping independently; then predict the consequence of this change: Multiply one feature by 1,000 and watch KNN/SVM geometry change before and after scaling.
Practitioner lens
Use it responsibly
Fit scaling only on training data and include it inside a pipeline. Inspect transformed distributions and remember that scaling does not fix outliers or skew by itself.
Transfer testTransfer this idea to a new example and justify each decision using this practitioner rule: Fit scaling only on training data and include it inside a pipeline. Inspect transformed distributions and remember that scaling does not fix outliers or skew by itself. Then explain what should change if you deliberately test: Multiply one feature by 1,000 and watch KNN/SVM geometry change before and after scaling.
Worked exploration
Use the visual as an experiment, not decoration
Compare Euclidean distance using age (years) and income (dollars). Without scaling, income dominates because its numeric range is much larger. Standardise both features and recompute the nearest neighbour.
Technical lens
Standardisation centres and rescales by training mean/std; Min–Max scaling maps a range; robust scaling uses resistant statistics. Tree splits generally do not require scaling.
Practitioner check
Fit scaling only on training data and include it inside a pipeline. Inspect transformed distributions and remember that scaling does not fix outliers or skew by itself.
Prediction before interaction
Multiply one feature by 1,000 and watch KNN/SVM geometry change before and after scaling.
Exploration walkthrough
Turn the interaction into an evidence trail
Compare Euclidean distance using age (years) and income (dollars). Without scaling, income dominates because its numeric range is much larger. Standardise both features and recompute the nearest neighbour. Before moving the control, state your prediction. After the visual changes, name the specific state, statistic, boundary or mapping that changed and explain why that change is consistent—or inconsistent—with your prediction.
Record one observable quantity before the interaction and the same quantity afterwards.
Change one factor at a time so the causal effect of the control is inspectable.
Use an edge or failure case to discover where the concept stops behaving as the simple story suggests.
Static orientation diagram for Scaling & Distance; use the interactive visual above to test how the relationships change.
Reference depth
Open the complete material
The flagship experience is the map. These pages contain the roads.
