Learning Hidden Structure from Data: A Method for Marginalizing JointDistributions Using Minimum Cross-Correlation Error

摘要

This thesis demonstrates an entropy-based, Bayesian dependency algorithm-MinimumError Tree Decomposition II (METD2)-that performs computer-based generation of probabilistic hidden-structure domain models from a database of cases. The system learns probabilistic hidden-structure domain models from data, which partially automates the task of expert system construction and the task of scientific discovery. Existing probabilistic systems find associations among the observable variables but do not consider the presence of hidden variables, or else, do not use cross-correlation error as the metric for building the hidden structure. The algorithm decomposes a joint distribution of n observable variables into n+l observable and hidden variables. The hidden variable exists in the form of a tree consisting of n-l interior nodes. The final product of the procedure is a combined tree whose n leaves are the observable variables in a sample and whose n-l interior nodes are the marginalizations for the leaves.