Leaming Influence Diagram from Data
Proceedings of the Sixth International Workshop on Artificial Intelligence and Statistics, PMLR R1:183-190, 1997.
There are many cases where decisions are made (and actions are taken) repeatedly under uncertainty, and consequences (results) ofthose decisions are available. For example, in telecommunications industry repeatedly decisions are made every day for fraud detection and account treatment. Indicators (variables) that have large uncertainties are used to make these decisions. Furthermore, the consequences of such decisions are recorded for later analysis. Similarly, in the financial industry, stocks or currencies are traded based on some indicators (variables). The consequences of these trade can be found. Similarly in the medicine, the patient treatment decisions are made on the basis of the patient information, and the consequences of these decisions to the patients can be found. These data sets contain uncertain variables, decision variables, and value lottery (final outcomes). Furthermore these decisions may not be made not by a single decision maker, but by many decision makers. In contrast to a typical decision analysis, in these environments decisions are made repeatedly. This paper addresses the discovery of knowledge bearing on these decisions in the form of influence diagrams (normative decision models) using a novel supervised machine learning method that constructs Bayesian network models with decisions. Algorithms presented in this paper exploit the goal oriented characteristics o f influence diagrams and generate a specific form of influence diagrams that are efficient, both to learn and evaluate. For this reason they are called "efficient" influence diagrams.