Building Tables from Other Nets

There are many ways that Netica can learn the CPTs of nodes.  It learn from case data, learn from a case file, learn from an Excel file or learn from other nets (called source nets).   In the latter case, you would be building a destination net that represents the same world as a source net.   In other words, Netica will learn the CPT tables of a sub-net within a source net.  All the inference results will be the same in both nets.  The difference is that it is using a different representation, so the link structure can be different.  In particular, nodes that have links in one net, may not have links in the other and the link directions might be different.

How to: open an existing net, or create a new one.  If it is a new net, you may want to copy and paste nodes from the source net into the destination net to ensure the node and state  names are the same.  You can then select a single node, or no nodes (in which case all nodes will be learned).  Next choose Table Build From Other Net. Netica will automatically learn the sub-set of CPTs from the source net.

Notes:  Nodes in the destination net must  have the same name and same state names as the source net, but they can have different titles.  You may want to copy & paste nodes between nets (and delete the resultant input links) to achieve this.

Netica can't learn CPTs from the source net if there is a parent node in the destination net that is not in the source net.  To resolve this,  first learn the tables from the source net, then add the new parent(s) to the destination net.  This is not true for children nodes.

In both cases, Netica will learn as many of the nodes as possible.

Applications:  when you are learning a net from data,  the link structure used can significantly affect the quality of learning.  For example, if you have a target node, Netica can determine a very effective TAN structure.  However, to combine with other sub-nets, you may want a different link structure.

So, you can create a net with the link structure you want and then build it's CPTs using the learned net as the source net.