ReseauLu Approach

Using the ReseauLu approach the data is represented as a set of graphical objects (nodes) into a two dimensional space. Each node is characterised by 2 aspects: its properties and its links.

- Properties: in our example, its size is linked to its observed frequency, and its shape depends on its structural position.

- Links: The association matrix or contingency table provides the basis for the visualisation of links. Only the most specific couples are selected (i.e. cells of the association matrix having the highest values of the normalized difference between observed and expected values).

There are 3 main stages using different algorithms: (i) selection of nodes, (ii) constructing association matrixes and selection of relevant links, (iii) mapping of specific relations.

1- Selection of nodes
The vector of nodes is sorted by number of inlinks in descendent order.
Selection of top 50 (top 100) means that the 50 most cited nodes is selected.

2- The construction of association matrixes
The starting point is the simple cross tabulation of data collected. This data is transformed in 3 steps: (i) weighted matrix, (ii) construction of matrix of expected values, and (iii) association matrix.

Step 1 is optional. It enables to take into account specific criteria characterising items observed. The software offers to associate to any item a vector of numerical properties.

Step 2 builds the matrix of the expected values corresponding to the "0" hypothesis of the complete statistical independence of the rows and columns of the table. The value of each cell is replaced by the combination of corresponding marginal values (i.e. multiplication of the totals of the corresponding raw and column divided by the general total).

Step 3 builds the association matrix. It calculates the normalized difference between observed and expected values using the following formula (O-X)/SQR(X) where O is the observed value and X, the expected one.

Traditionally, the most positive values translate the most specific associations between items. However there are strong limitations to such manual interpretation, this explains the software has a specific mapping interface.

2- The mapping of specific relations

The algorithm optimises the positioning of objects in a two dimensional space focusing on the existence of "strong" ties. The initial binary matrix of links can be represented without deformation only in a multidimensional space. To minimize the deformation of the final map in a two dimensional space, the software uses a dynamic positioning simulating the interaction between objects. It does so through a three step optimisation process: (i) global initial positioning of the object vis-à-vis all the other objects in the space; (ii) micro-optimisation of the positioning of the object vis-à-vis the other objects to which it is directly connected ("network neighbours"); and (iii) meso-optimisation of groups of highly connected objects ("clusters"). The optimization process depends on explicit rules defining symmetry properties, structural equivalence of points inside the structure, centrality and "betweeness" of objects.

One of the interests of the method is to adapt the degree of specificity to the level of analysis. A strong degree of specificity (say 30%) enables to develop global visions of the relative positioning of countries in the space of sectors. Increasing the cutoff level (say to 40%) enables to learn more on a pole, while a cutting level of 50% enables to visualise the full specificity of an object.

Workshop locations