🧮 Data Preprocessing Hyperparameters

About this simulation

Before a model ever sees training data, preprocessing choices reshape it. This scene treats each sample as a point in 3D feature space (X, Y, Z) and lets you change the scaling method — a genuine preprocessing hyperparameter — along with how skewed the raw distribution is and how many outliers it contains.

How it works

• None (raw) — points sit at their original, unscaled values, so features with large natural ranges dominate the cloud's spread.
• Min-Max — every feature is linearly compressed into a fixed box; a single outlier can squash the rest of the data into a sliver near one wall.
• Standardization (z-score) — features are centred on 0 with unit variance; still sensitive to outliers because both mean and std are outlier-hungry statistics.
• Robust scaling — centring uses the median and spread uses the interquartile range, so the bulk of the cloud stays stable even as outlier share climbs.

Controls

• Scaling method — pick which transform is applied live to every point.
• Outlier share — fraction of samples pushed far from the core cluster (red points).
• Feature skew — skews the X feature's raw distribution toward one tail before scaling.
• Sample count — number of points in the cloud.
• ↺ Resample data — draws a fresh random dataset with the current settings. Drag to rotate, scroll to zoom.

Did you know?

Min-Max scaling and standardization are both linear transforms, but because they use different reference statistics (min/max vs mean/std), an extreme outlier can distort them very differently — which is exactly why robust scaling exists as a preprocessing hyperparameter choice for outlier-heavy datasets.