Rule optimization for classification and detection

Abstract

This disclosure describes methods, systems, and computer-program products for determining classification rules to use within a fraud detection system The classification rules are determined by accessing distributional data representing a distribution of historical transactional events over a multivariate observational sample space defined with respect to multiple transactional variables. Each of the transactional events is represented by data with respect to each of the variables, and the distributional data is organized with respect to multi-dimensional subspaces of the sample space. A classification rule that references at least one of the subspaces is accessed, and the rule is modified using local optimization applied using the distributional data. A pending transaction is classified based on the modified classification rule and the transactional data.