Decision Support Beyond Decision Making: from Model Extraction to Explanation

Within an organization, different types of decisions have to be made on a daily basis ranging from deciding whether to invest in a new factory to deciding whether a person is eligible for a loan. Typically, three types of decisions are identified namely strategic (focus on the long term), tactical (focus on the medium-term) and operational (focus on the short term). Operational decisions have a high impact on companies because of the fact that these are taken regularly on a daily basis which creates a high volume of decisions which influence the organizations. This doctoral thesis investigates how the decision support and management of operational decisions can be supported on multiple aspects. In order to support operational decisions, the Decision Model and Notation (DMN) standard can be used to effectively model, communicate, execute and even automate operational decisions. In short, DMN models contain both the decision structure and the decision logic of an operational decision. The DMN standard is chosen as a way to model operational decisions in this thesis given its use in multiple knowledge-intensive organizations such as financial institutions, governments and health organizations. This doctoral thesis first investigates the feasibility to extract DMN models from textual descriptions using NLP models. Both encoder (BERT) and decoder (GPT-3, GPT-4) NLP models are used to extract DMN models from textual descriptions. It can be concluded that these models seem to be able to extract the decision structure from a textual description but struggle more with extracting the decision logic due to having difficulties with correctly identifying what the relevant decision domain is. Secondly, the automatic generation of advice and explanation of decisions using DMN models is investigated. To do this, DMN chatbots are studied and enhanced with NLP models as intent interpreters to identify how the DMN chatbot can answer the questions of a user. Lastly, thanks to information systems it is possible to track and store each step in a decision process in event logs. By considering multiple facets of a decision process such as the customer and the product multiple objects are involved which are then stored in object-centric event logs. In the last part, a novel object-centric event log format is developed allowing to track attribute changes in a decision process. Next, an algorithm is developed to extract DMN models from object-centric event logs allowing to compare the implemented decisions with the prescribed decisions in the textual decisions. Finally, a framework is developed comparing multiple object-centric event log formats with their respective advantages and disadvantages.