Entropy-Based Learning of Compositional Models from Data

We investigate learning of belief function compositional models from data using information content and mutual information based on two different definitions of entropy proposed by Jirousek and Shenoy in 2018 and 2020, respectively. The data consists of 2,310 randomly generated basic assignments of 26 binary variables from a pairwise consistent and decomposable compositional model. We describe results achieved by three simple greedy algorithms for constructing compositional models from the randomly generated low-dimensional basic assignments.