The Chernoff Faces method is a data visualization technique brought to us by the 70's. It was developed by Herman Chernoff to represent multivariate data, ostensibly effectively representing up to 18 variables. Facial features(eyes, nose, eyebrows) are mapped to multiple variables, with size, orientation, shape, color, and placement potentially representing different attributes of a single observation.
Looking for more info? Check out our presentation!
The Chernoff faces technique is an interesting way to represent multivariate data. It can be used to detect similarities between different items, but it is not the most efficient or the most accurate way to do so. Other techniques, such as parallel coordinates, star graphs, or radar charts, depict as many dimensions as Chernoff faces, but are easier to interpret.
Implementing visualizations of Chernoff faces is quite challenging. Data must be normalized, and often binned in order to convey meaning. Scaling and normalization can be complicated, particulary if variables represent many different kinds of data. On the viewers end, it becomes difficult to extrapolate meaningful quantitative data from normalized and binned representations.
Faces do not make it easy to present data without inate perceived bias. For example, a curved mouth holds positive and negative connotations, which must be considered in order to avoid unwanted implications.
Our implentation uses a Chernoff face plugin for D3 designed by Jason Davies. We normalized our data and mapped to defined paths representing facial attributes.
Each face is a representation of a school. The face size represents acceptance rate, the nose width represents founding year, the nose height represents size, the eye width represents rank, hair represents the founding affiliation, and the mouth represents the state. We ran into issues normalizing categorical data appropriately, so certain attributes appear binary, when in fact they have more levels.
Midd is still pretty cute.