Definition of the Key Concepts
The following is an overview of the concept for a knowledge-based cause-effect analysis. A detailed description can be found in Benjamins (2014). The concept uses the separation of the knowledge representation and the knowledge reasoning, which is typical for a knowledge-based system (Brachman et al. 2004).
For the knowledge representation, knowledge about cause-effect relationships is extracted from a variety of sources within a company, transformed into a unified structure, and loaded into the knowledge base. The unified structure consists of knowledge elements, relationships between these elements, and different relationship types. The sources can contain structured and unstructured knowledge. Most importantly, formulas of DSS models provide structured knowledge about quantified relationships between data. OLAP cubes also implicate potential cause-effect relationships within their dimensional structure and even ETL processes from a data warehouse (DW) implicitly contain cause-effect knowledge within their transformations. Linked (open) data from the World Wide Web can add external influences. This is enriched with unstructured knowledge about cause-effect relationships from experts to add relationships between data from different sources. The result is a homogeneous knowledge base with knowledge about cause-effect relationships from a variety of sources (Figure 7.1).
The knowledge reasoning is applied to the knowledge base in order to identify relevant causes in a specific decision-making situation. The reasoning is based on the separation of a decision-making process into the three phases: intelligence design, choice (Simon 1977), and the human approach to solving problems by using
Figure 7.1 Overview of a knowledge-based cause-effect analysis. (Adapted from Benjamins, A., Knowledge-based cause-effect analysis for context-driven decision support. In DSS 2.0—Supporting Decision Making with New Technologies, Supplemental Proceedings. Paris: IFIP Working Group 8.3 digital publications, http://dss20conference.files.wordpress.com/2014/05/benjamins.pdf, 2014.)
a specific problem context (Newell and Simon 1972). The reasoning is divided into three phases with two steps in each (Figure 7.1). During the initialization phase, a decision-making situation is isolated and matched into the knowledge base. This identifies the relevant factors for the decision-making situation within the knowledge base. The exploratory phase uses these identified factors in the activation and validation steps. During activation, all relationships between elements directly connected to the identified factors are marked as unconfirmed cause-effect relationships. These unconfirmed relationships are statistically validated with the help of time series data, for example, from a DW or operational systems. If they are confirmed, the connected elements are promoted to relevant factors and used for another activation step. The result is a targeted activation along a chain of promising factors in a specific decision-making situation. The evaluation phase offers the user the possibility to verify the confirmed cause-effect relationships and, if necessary, to adjust the affected relationships in the knowledge base.