Controlling Some Statistical Properties of Business Rules Programs

Abstract

Business Rules programs encode decision-making processes using "if-then" constructs in a way that is easy for non-programmers to manipulate. A common example is the process of automatic validation of a loan request for a bank. The decision process is defined by bank managers relying on the bank strategy and their own experience. Bank-side, such processes are often required to meet goals of a statistical nature, such as having at most some given percentage of rejected loans, or having the distribution of requests that are accepted, rejected, and flagged for examination by a bank manager be as uniform as possible. We propose a mathematical programming-based formulation for the cases where the goals involve constraining or comparing values from the quantized output distribution. We then examine a simulation for the specific goals of (1) a max percentage for a given output interval and (2) an almost uniform distribution of the quantized output. The proposed methodology rests on solving mathematical programs encoding a statistically supervised machine learning process where known labels are an encoding of the required distribution.

Universality and prediction in business rules

“…That is a problem that industrial BR systems would be very interested in. In the work of Wang and Liberti, 49 we look at discrete distributions, for example, but many other possibilities exist.…”