From Business Goals to Analytics and Machine Learning Solutions: A Conceptual Modeling Framework

Machine learning and analytical techniques are increasingly becoming an integral part of many types of information systems. Despite the hype around these technologies, modern organizations continue to struggle with using them to take advantage of their data and address business problems. The development of business analytics and machine learning solutions in enterprises involves tackling a wide range of complexities with respect to requirements elicitation, design, development and deployment of such solutions. In spite of the necessity and relevance of conceptual modeling and requirements engineering approaches to the process, not much research has been done in this area. This thesis develops, demonstrates and evaluates a conceptual modeling framework for business analytics and machine learning solutions. The framework includes meta-models, methods, design catalogues and patterns, guidelines, and instantiations. It consists of three modeling views, representing different aspects of a solution and viewpoints of different roles involved in the development of such systems. The Business View supports the elicitation of business analytical requirements by capturing stakeholders, strategic goals, decisions, questions and required insights. The Analytics Design View supports the design of the solution by capturing algorithms, metrics, and quality requirements, focusing primarily on machine learning solutions. The Data Preparation view supports the design of transformation workflows by capturing data tables, flows, and preparation tasks and operations. These views are linked together to represent a holistic view and bridge the gap from business strategies to machine learning algorithms to data preparation operations. The framework comes with a set of design catalogues and solution patterns that encode and represent generic and well-proven machine learning solutions for commonly-known recurring business problems. The framework is also augmented with methodological steps and modeling guidelines for supporting domain users in working with the framework. It also includes model-based support for linking analytics-driven insights to consequent enterprise actions and changes. The framework is tested through multiple real-world projects, an empirical study in the healthcare domain, and prototype design and implementation in collaboration with a large information technology vendor. This thesis reports on findings from evaluating the expressiveness and usefulness of the framework, in addition to revealed drawbacks and improvement opportunities.Ph.D