| Knowledge discovery in databases (KDD) and data mining systems in particular are intended to serve as decision support systems, now business intelligence systems. Yet, current data mining methodology uses knowledge discovery techniques to predict the likelihood of specific events, or describe likely patterns (clusters or classes) using only the information residing in extant databases, data or text. The discovered knowledge is assumed to be interpretable and useful for decision-making purposes. In reality, the output interpreted by data mining experts---who have limited understanding of the business environment, business objectives and constraints, managerial decision preferences and options---using only objective measures of assessment. Data mining results are regularly irrelevant or uninteresting to the decision maker.;This research proposes an approach for eliciting, and incorporating decision maker prior knowledge into the data mining process. Using a multicriteria decision analysis approach and rule-based systems, we present an approach that inquires into decision maker prior knowledge and re-expresses it in machine-readable forms that can be integrated with database information for data mining and better decision support. By using multicriteria decision analysis we can capture and incorporate decision objectives, constraints, decision preferences, uncertainty, as well as decision alternatives associated with domain problem under consideration. Furthermore, we evaluate the resulting data patterns using both technical (objective) and subjective measures of interestingness.;Our approach is formulated as method in the tradition of design science. It is instantiated an evaluated within the practical domain of the management and prognosis of traumatic brain injury patients. We evaluate the data mining results using our approach against current methodology, and the findings of our study show that for domains similar to our case study domain, the incorporation of prior domain knowledge into the data mining process improves pattern interestingness from both objective and subjective assessment. We also found that prior knowledge specificity was an important moderating factor, the higher the degree of specificity of the articulated prior knowledge the more discernible the effect on pattern interestingness. Our method is however more computationally expensive. In the practical world, and after twenty tow cycles of Moore's Law, the cost may be worth bearing if real knowledge discovery is the pay off. |
