Baseball / Statistics

What's the Matter With Chernoff Faces?

8 August 2006

In 1976 statistician Herman Chernoff had the idea of representing multivariate data in figures now known as Chernoff Faces. The theory is that since we are highly practiced in the art of facial recognition, and can discern minute variations in features and expression, perhaps encoding data in a likeness of a human face would reveal things that, say, a bar graph wouldn't. And reading the diagram would, ideally, be as natural as interpreting the face of your best friend. As an example, here are some team statistics from the 2005 baseball season represented in a table and then as a series of Chernoff Faces:

win pct

face height, smile curve, hair styling

hits

face width, eye height, nose height

home runs

face shape, eye width, nose width

walks

mouth height, hair height, ear width

stolen bases

mouth width, hair width, ear height

For me, the Chernoff Faces have never been particularly effective or easy to read. While I can effortlessly tell how Pedro Martinez is feeling on the mound by looking at his face (on the TV broadcast), when reading Chernoff's faces I have to stop and think carefully about each one. It's not natural or easy, and usually I'd prefer a good old fashioned bar graph.

Specifically, what's wrong with Chernoff Faces? Allow me to propose that the problem lies in the un-human-ness of the stick figure drawings. They look more like bits of a breakfast cereal than a person. What's needed here are realistic human faces. And so, I humbly submit, for your consideration, Reisner faces:

win pct

cap rotation

hits

mouth shape

home runs

tongue length

walks

left eye size

stolen bases

right eye size

Since, clearly, a more readable expression for the Mets in 2005 would be the one at right it seems I haven't done any better than Chernoff.

Oh well.