Chernoff Face — Definition, Formula & Example
A Chernoff face is a chart that draws each data point as a cartoon face. Each variable controls one facial feature, such as mouth curve, eye size, or face width. Similar data points end up with similar-looking faces.
Herman Chernoff introduced the technique in 1973. Each observation in a -variable data set becomes one face glyph. Each variable is assigned to a facial parameter , such as head shape, eye size, nose length, or mouth curvature. A monotone function, usually linear, maps the observed range of onto the allowed range of . Chernoff's original design allowed up to 18 facial parameters.
Key Formula
Where:
- = The observation's value of the variable
- = The smallest and largest values of that variable in the data set
- = The smallest and largest allowed settings of the facial feature
- = The feature setting used to draw this observation's face
Worked Example
Problem: A used-car data set has fuel economy from 20 to 40 mpg, price from $10,000 to $30,000, and age from 0 to 12 years. You map the variables to features as follows. Fuel economy sets mouth curvature, from −1 (full frown) to +1 (full smile). Price sets face width, from 2 to 4 inches. Age sets eye height, from 0.2 to 0.6 inches. Find the feature settings for a car that gets 35 mpg, costs $15,000, and is 9 years old.
Step 1: Rescale fuel economy to mouth curvature. Here 35 mpg is 75% of the way from 20 to 40.
Step 2: Rescale price to face width. A price of $15,000 is 25% of the way across the price range.
Step 3: Rescale age to eye height. An age of 9 years is 75% of the way from 0 to 12.
Answer: This car's face has a mild smile (curvature 0.5), a narrow face (2.5 inches wide), and fairly large eyes (0.5 inch tall). A cheap, newer, gas-guzzling car would instead get a frown, a narrow face, and small eyes.
Why It Matters
In a multivariate statistics or data visualization course, you often need to compare dozens of observations on five or more variables at once. A scatter plot shows only two or three variables by position. People are very good at spotting small differences between faces, so a grid of faces can reveal clusters and outliers that a data table hides.
Common Mistakes
Mistake: Assuming the choice of which variable controls which feature does not matter.
Correction: Viewers notice the mouth and eyes far more than features like ear size or nose length. Swapping assignments can change which observations look alike. Put your most important variables on the most noticeable features, and state your mapping in a legend.
