Automated Decision Modeling with DMN and BPMN: A Model Ensemble Approach

Abstract

Plethora of available heterogeneous transactional data and recent advancements in machine learning are the key forces that enable the development of complex algorithms that can reach human-level performance on an increasing number of tasks. Given the non-linear structure composed of many layers of computation, these highly accurate models are usually applied in a black-box manner: without a deeper understanding of their inner mechanisms. This hinders the transparency of the decision-making process and can often incorporate hidden decision biases which are potentially present in the data. We propose a framework for generating decision-making models conforming to Decision Model & Notation standard based on complexity-reducing techniques. An ensemble of decision-tree classifiers in a layered architecture is proposed to control the bias-variance trade-off. We have evaluated the performance of the proposed method on several publicly available data-sets tightly related to socially sensitive decision-making.