For relational datasets, each dataset contains THREE files: a data file, a link file, and a relational template.

The format of the data file is one example per line, and the format for a line is:

<example_id> type subpopid label feature1 feature2 ...

The format of the link file is:

<ID of starting example> <ID of to example> <link type>

The format of the relational template file is:

<Operation> < link type>

where

• type=b or k (for binary or ordinary examples)
• subpopid is NUL or a string, naming the subpopulation from which the example was drawn
• label is +1 or -1 for binary
• features are a feature name (for binary features) and a featureName=value for numeric features.
• Operation are either ‘EXISTS’ or ‘COUNT’

Here’s an example dataset.

In minorthird, there’s a test case for stacked graphical learning. You can refer to the webpage here to find out how to install minorthird. After installing minorthird, go to minorthird/demos/:

 
NumericDemo_SGM.java:
 Loads a relational dataset; trains and tests a stacked graphical model based on an MaxEnt learner on the given data.
 
After compiling, use the following command to run the test case
java NumericDemo_SGM SGMSample/all.localmodel SGMSample/ngb.lsf SGMSample/relTemp.txt