Validation Mode: Learning | Clustering | Variable Clustering
New Feature: Post-Processing of Leaf Nodes
For Bayesian networks with a Naive structure or sub-structure, i.e. a node that is parent of many conditionally independent children, the hierarchical variable clustering algorithm merge the parent node with the children that have the strongest connections with the parent. Once the strength of the connection too week to allow the clustering with the newly formed cluster, or when the maximum number of nodes per cluster is reached, the remaining children become virtually orphans and are not associated with any cluster.
In order to reduce the risk to have such single node clusters, we have extended our algorithm to simulate the duplication of the parent node for keeping the connection with the virtual orphans. The example below describes, step by step, how the clusters are constructed on a Naive structure, with and without this new option. The maximum number of nodes per cluster in this example is 5.
New Feature: Settings Dialog Box
Instead of using the parameters defined in Window | Preferences | Learning | Variable Clustering, this new dialog box allows to directly select the parameters for the clustering algorithm.