We were recently introduced to an interesting network analysis application which could be tackled by ClustanGraphics.  The objective was to analyze traffic between the nodes on a supply network with a view to establishing a regional organization clustered around regional distribution centers.

Our recommended approach involved treating the traffic between the nodes as a square similarity matrix.  This is read as proximities in square matrix format, and converted to a proximity matrix by totalling the traffic between any two nodes.

We recommended clustering the traffic values using Average Linkage, and selecting cluster levels in the range 10-20 clusters for closer examination.  For each cluster, ClustanGraphics identifies one node as the cluster's exemplar.  In this application, it is the node with the highest average traffic level with the other nodes in the cluster.  The cluster exemplars are then candidates for conversion to regional distribution centers, with the nodes in each cluster classified into network regions.

An alternative approach is to find a geographically optimized network.  Here we enter the grid co-ordinates of the location of each node, and weight the nodes by their total traffic volume.  Using Direct Data Clustering in ClustanGraphics, the nodes are clustered into geographical regions to minimize the sum of the distances between each node and the cluster center, weighted by the traffic at each node.  The nodes which are closest to the weighted cluster centers are then candidates for regional distribution centers, thereby minimizing the traffic between regional centers and the nodes in each region.