Budgeted parameter learning of generative Bayesian networks.

摘要

This dissertation studies the parameter estimation problem of Bayesian networks in the budgeted learning setting. More precisely, we assume that the correct structure of the Bayesian network representing the underlying distribution is given together with a fixed positive budget, and each data attribute of the training set is associated with a cost. During the training phase, the learner is allowed to purchase value of an attribute of a certain data instance by deducting the corresponding cost from the budget. The goal of the learner is to make the purchases wisely so that when the budget is exhausted, the learned parameters from the purchased data are as close as possible to the underlying distribution that generates the data. The dissertation presents a theoretical framework for the problem, analyzes its hardness, and compares different algorithms and heuristics for solving the problem efficiently and economically.