The decision exploration lab: supporting the business analyst in understanding automated decisions

Bertjan Broeksema

Research output: ThesisThesis fully internal (DIV)

830 Downloads (Pure)

Abstract

A Decision Management System (DMS) provides means to model and automate enterprise decisions and they are applied in a wide range of industries, among which health care, commerce, insurance, finance and transportation. These systems make millions of decisions each day without direct human supervision, impacting the life of millions of people and impacting economies at a large scale. The multiplicative effect of decision automation provides the opportunity to
fine-tune the decision system. By analyzing its global and emerging properties rather than focusing on the details of each decision, the system as a whole can be better adapted to the reality it models.

Like expert systems, DMSs provide a clear separation of decision logic, information related to individual decisions and decision execution. These data spaces contain a wealth of information related to the structure and functioning of a DMS. In this thesis various ways are explored to visualize and analyze this data in order to help a business user to gain a deeper understanding of automated decisions.

To address the problem of understanding the global and emerging properties of automated decision making systems, we combine interactive analysis of the decision data with analysis of the decision logic. We present a visual analytics system, the Decision Exploration Lab (DEL), which provides a verbal analysis mode and a visual decision exploration mode. In verbal mode the user can make selections on past decisions using controlled natural language. In visual de-
cision exploration mode, the decision data is analyzed using Multiple Correspondence Analysis (MCA). The analysis results are visualized using interactive techniques to show the important structure of the decision data to the user. Correlated concepts can be clustered at a level of granularity that suits the needs of the business analyst. Clustered concepts can next be linked to the
rules of the decision logic that are relevant for the subset of decisions which match these concepts. We evaluated our approach with two use case scenarios from the car insurance industry. Apart from the above, we propose a number of technical contributions, enhancements and extensions to information visualization methods, for multivariate categorical data. Firstly, we
present a generic algorithm to generate all well-known treemap layouts as well as other rectangular space-filling layouts. Secondly, we present explanatory and interactive visualization techniques to support interpretation and usage of MCA. Thirdly, we present labeling and scale adjustment techniques in order to improve the usability of 2D-plots.
Original languageEnglish
QualificationDoctor of Philosophy
Awarding Institution
  • University of Groningen
Supervisors/Advisors
  • Telea, Alexandru, Supervisor
  • Melançon, Guy , Supervisor, External person
  • Baudel, Thomas, Supervisor, External person
  • Fekete, Jean-Daniel, Assessment committee, External person
  • Kerren, Andreas, Assessment committee, External person
  • McGuffin, Michael J., Assessment committee, External person
Award date3-Mar-2014
Place of Publication[S.l.]
Publisher
Print ISBNs9789036767866
Electronic ISBNs9789036767859
Publication statusPublished - 2014

Cite this