Each axis needs independent normalization

What it is

Parallel coordinates require independent scaling and normalization for each dimension to map different units and ranges onto a common display height. This is a display transformation, not a claim that raw values have become commensurable. Equal height means equal relative position under each axis mapping. Putting income, temperature, and ratings on one raw numeric range compresses smaller magnitudes, making polyline trajectories reflect unit choice rather than record profiles.

Why it happens

Linear min–max mapping preserves within-axis order and relative distance but depends on endpoints and outliers. Z scores express displacement from a center and scale while remaining sensitive to distribution and extremes. Quantile mapping makes equal height approximate equal rank but sacrifices raw spacing. Log transforms fit theoretically meaningful multiplicative scales, with separate treatment for zero, negatives, and unit interpretation. Reversing an axis changes segment texture without changing its data and should be recorded separately from numerical transformation.

Studying it

Compare linear range, robust range, z-score, quantile, and justified log mappings on anomaly, record-matching, association-direction, and business-threshold tasks. Manipulate outliers, skew, unit conversion, drift, and inversion; measure accuracy, time, and interpretations of equal height. For cross-chart comparison, test per-view adaptive ranges separately from shared reference ranges. They support different tasks and should not be judged only by how fully marks occupy the display.

Where it stops holding

Not every dimension should receive the same normalization. Preserve real ticks when absolute thresholds matter; bounded scores, cyclic quantities, nominal categories, and diverging signed values need semantic axes. Quantile mapping draws unequal raw gaps as equal rank gaps and cannot support effect-size reading. Adaptive ranges reveal within-view variation but make equal height differ across charts; shared ranges support comparison but may be exceeded under drift. Missing values have no numeric position and must not enter normalization as zero.

Applying it

• Show original unit, observed range, display range, transformation, and direction on each axis, with keyboard and screen-reader access to raw values.
• Select scales by task: linear or robust mappings for profiles, quantiles for rank, and logs for multiplicative meaning. Never silently clip or alter values.
• Mark reversed axes continuously with high/low labels and state, and allow restoration; do not hide inversion merely to reduce crossings.
• Use a versioned reference range across compared views. If each view adapts independently, explicitly disallow height comparison across views.
• Test unit invariance and edge cases: valid unit conversion should preserve normalized position, while zero, negatives, outliers, missingness, and new out-of-range values require explicit outcomes.

Related

• Same group: U2.14.1 Parallel coordinates render each record as a polyline across axes · U2.14.2 Only adjacent-axis relations read directly · U2.14.3 Axis order decides which patterns can be found · U2.14.4 More records, more overplotting, less structure
• Nearby: U1.11.3 Only charts sharing a scale can be compared on magnitude · U3.03.1 Log scales suit data spanning orders of magnitude
• Search terms: per-axis scaling · min-max normalization · quantile transform · axis inversion