Notes
🔍

Coefficient of Variation

The same amount of “spread” can mean very different things depending on the scale you’re measuring. A $5 variation is huge for a $20 item, but tiny for a $2,000 item. The coefficient of variation is a simple way to express variability in a scale-free, comparable way.

What it is
The coefficient of variation (CV) measures relative dispersion: it’s the standard deviation divided by the mean. It’s often reported as a percentage:

CV = (standard deviation / mean) × 100%

A higher CV means the data vary a lot compared to their typical level; a lower CV means values cluster more tightly around the mean.

Why it’s useful
Because CV is unitless, it lets you compare variability across different units or scales—something raw standard deviation can’t do fairly.

  • House prices: Neighborhood A has SD = $50k, mean = $500k → CV = 10%. Neighborhood B has SD = $50k, mean = $250k → CV = 20%. Same SD, but B is relatively more volatile.
  • Exam scores: A test with mean 80 and SD 8 has CV 10%; another with mean 40 and SD 8 has CV 20%—scores are less consistent relative to the average.

Where it shows up in AI/ML
In ML, CV helps during exploratory data analysis and feature screening:

  • Detecting near-constant features (very low CV) that may add little signal.
  • Comparing variability across features with different units (e.g., income vs. age).
  • Monitoring stability of metrics or sensor signals over time (high CV can indicate drift or noise).

Important caveats
CV can misbehave when the mean is near zero (it can explode) or when data include negative values where “relative to the mean” becomes hard to interpret. It’s most meaningful for ratio-scale, strictly positive measurements.

Coefficient of Variation (CV) is a unitless measure of relative variability, defined as the standard deviation divided by the mean (often expressed as a percentage). It enables dispersion comparisons across features with different scales or units, which is useful in ML for assessing feature stability and noise. Example: if feature A has mean 100 and SD 5 (CV=5%) and feature B has mean 1 and SD 0.2 (CV=20%), B varies more relative to its level.

Imagine two jars of candy. In Jar A, the number of candies changes by about 5 each day; in Jar B, it also changes by about 5. If Jar A usually has 20 candies and Jar B usually has 200, that same “5” is a much bigger deal for Jar A. The Coefficient of Variation captures this idea: it measures how spread out the data is relative to its typical size (its average).

In statistics and machine learning, it’s useful for comparing variability across things with different scales, like prices vs. temperatures, or model errors across different datasets.