Bertin's Classifier

Background

Jacques Bertin was the director of the cartographic laboratory at the École à Pratique des Hautes Études who specialized in the display of geographic information. He wrote several of the classic texts on this topic, including in particular Semiologie Graphique: Les Diagrammes, Les Reseaux, Les Cartes (Mouton: Paris, 1967), one of the first systematic examinations of the art and science of map-making.

Classification

But Bertin was not content with map-making, nor statistical graphics, and not even with the static graphic representation of information. In a further classic, La Graphique et le traitement graphique de l'information (Flammarion: Paris, 1977) he advocated the use of graphics in data exploration.

An example

On p.33 he introduces the problem of classifying communities of varying sizes in France. He begins by noting nine properties together with their realization (or lack thereof) in sixteen villages and towns in France.

The properties he noted were the following:

  1. College presence in the town or village of a high school
  2. Cooperative Agriculture, presence of an agricultural cooperative
  3. Gare, presence of a train station
  4. École Classe Unique single-room elementary school
  5. Veterinaire veterinary
  6. Pas de Medecin lack of a local doctor
  7. Pas d'Adduction d'Eau lack of running water
  8. Gendarmerie presence of a police station
  9. Remembrement whether the town or village has been

Although he doesn't provide information about the sixteen communities (in the columns A - P in the graphic), it is safe to say that it is difficult to recognize a pattern in them. He sets himself the task of classifying these communities along "natural" lines.

The Classifier

The mechanical classifier has only two basic operations, but both of them are logically complex. The first operation "shuffles" rows, aiming for an ordering in which similar rows are adjacent. The second operation shuffles like columns, again with an aim toward a result in which similar columns are next to each other. We first examine the effect of reshuffling the rows.

Notice what's happened here: the rows -- which, after all, occur in arbitrary order, are simply reordered. In (3) the original row numbers are retained so that you can examine the effect of the reordering more exactly. In the first three rows of (3), one can see, for example, that the properties 'high school', 'train station' and 'police station' seem to hold together of an individual community. That is, they're either all present or none of them is in any given community. With the exception of community C, which has only a police statioin, it turns out that only H and K have any of these facilities (in this data set), and they both have all of them. We can summarize: the first step has identified similar properties.

Next, the same procedure is applied to the columns. They likewise occur in arbitrary order, corresponding to the communities in the original data set, which we might view in any order. Reshuffling the columns, therefore, has the effect of grouping similar communities -- completing the classification task which is the purpose of the exercise.

Interpreting the Results

Even if the classification is complete, the result needs to be examined to see if it is interpretable. In other words, we need to check whether the results correspond to a useful classification. Bertin does this in a final step, which merely labels the groups which his classifier has identified.

A Software Realization

Peter Kleiweg has implemented a web version of the classifier, including Bertin's example, but also the opportunity to define one's own data sets.

Questions for Reflection

  1. In the final result, we can identify the groups in the rows as well as in the columns. What in the process is responsible for this?
  2. The classification is not perfect. Community C has only one of the three "urban" properties, while A and B both have two of the three rural properties and all of the intermediate ones. Finally, most of the villages show only two of the three the rural properties. Does this suggest that the classifier isn't working properly, or do you suppose that some rough edges are inherent in the task of natural classification?
  3. Some properties are represented negatively, e.g., 'pas de medecin' and pas d'adduction d'eau'. Using positive versions of these properties instead of the negative ones would seem to provide the same information, but how well would the process work? Suggest ways in which the classifier could be made more robust, i.e., less dependent on the form in which information is provided.
  4. Bertin's choice of properties in this example was "fortunate" -- they were all relevant to defining interesting classes. Suppose he had included properties such as "has a street called 'Main Street' (Rue de Ville)", "has more women then men inhabitants", or "is west of Paris" or other properties that turn out to be less useful. What would these do to the results? And what adjustments might be made to the classifier to allow it to function even when irrelevant properties are part of the input?
  5. All of Bertin's properties in this example are binary, i.e., they either hold or do not hold of a given individual (community). But other properites seem to be graded, e.g., proximity to highway system, amount of traffic, number of factories, etc. Speculate on what might be done to accommodate someone who wished to see graded properties play a role in classification.
  6. Kleiweg's software realization of Bertin's classifier has no access to the visual presentation that Bertin uses. Instead it counts fields with the same values in the rows (or in the columns), and it later presents it's results visually. Is it fair then to call this a visual technique?
  7. Bertin suggests his classifier is no longer useful when data matrices (such as that in (1), which serves as input to the process) becomes larger than 120 X 120 (p.31). Kleiweg's software implementation also has a limited size. The software might be optimized to deal with larger structures, but there is probably a limit to what we can "see in a glance". Does this suggest that graphic communication must be limited to communicating only relatively simple information?


Return to Graphic Communications Home page.