Computer models are representations of problem environment that facilitate analysis with high computing power and representation capabilities. They can be either inferred from the data using data mining techniques or designed manually by experts according to their knowledge and experience. When models represent environments that change over time, they must be properly updated or periodically rebuilt to remain useful. The latter is required when changes in the modelled environment are substantial. When changes are slight, models can be merely adapted by revision. Model revision is a process that gathers knowledge about changes in the modelled environment and updates the model accordingly. When performed manually, this process is demanding, expensive and time consuming. However, it can be automated to some extent if current data about the modelled phenomena is available. Databased revision is a procedure of changing the model so as to better comply with new empirical data, but which at the same time keeps as much of the original contents as possible. In the following we describe the model revision principles in general and then focus on a solution for a specific type of models, the qualitative multi-attribute decision models as used in DEX methodology.
TopBackground
The task of data-driven revision is to adapt the initial model to accommodate for the new data, while at the same time making use of the background knowledge, which was used in the construction of the initial model. Revision is to be applied only when the changes of the modelled concepts are not substantial, that is, if we deal with concept drift (Tsymbal, 2004). If the changes of the modeled system are substantial, it is usually better to construct the model from scratch.
Depending on the field of research, procedures of this kind are most often referred to as knowledge refinement or theory revision. Most of research in this field is done on propositional rule bases (Ginsberg, 1989; Mahoney & Mooney, 1994; Yang, Parekh, Honavar & Dobbs, 1999; Carbonara & Sleeman, 1999) and Bayesian networks (Buntine, 1990; Ramachandran & Mooney, 1998), but many principles of these methods are shared with those of revision procedures for other knowledge representations, such as case-based reasoning systems (Kelbassa, 2003).
Multi-criteria decision models (MCDM) are models used in decision analysis (Clemen, 1996). Data-driven revision can be a valuable tool for the ones that are used for longer periods of time. In our previous work, we have developed revision methods (Žnidaršič & Bohanec, 2005; Žnidaršič, Bohanec & Zupan, 2006) for two types of MCDM models of DEX methodology (Bohanec & Rajkovič, 1990; Bohanec, 2003). An input to MCDM models is criteria-based description of alternatives, where a model represents a utility function to evaluate the given set of criteria values. MCDM models are used for evaluation and analysis of decision alternatives. In contrast to traditional numerical MCDM (Saaty, 1980; Keeney & Raiffa, 1993; Triantaphyllou, 2000), the models of DEX methodology are qualitative and have utility functions in form of if-then rules. The concepts in these models are structured hierarchically and their values are defined according to the values of their immediate descendants in the hierarchy (see Figure 1). This dependency is specified with qualitative rule-based utility functions, which can be defined as crisp or probabilistic and are usually represented in tabular form (see Table 1). The concepts at the bottom of hierarchy serve as inputs, represent the criteria-based description of alternatives and must be provided by the user.
Figure 1. A simple hierarchy of concepts of a decision model for car purchase