Knowledge discovery in databases with joint decision outcomes: A decision-tree induction approach.

Abstract

Inductive symbolic learning algorithms have been used successfully over the years to build knowledge-based systems. One of these, a decision-tree induction algorithm, has formed the central component in several commercial packages because of its particular efficiency, simplicity, and popularity. However, the decision-tree induction algorithms developed thus far are limited to domains where each decision instance's outcome belongs to only a single decision outcome class. Their goal is merely to specify the properties necessary to distinguish instances pertaining to different decision outcome classes. These algorithms are not readily applicable to many challenging new types of applications in which decision instances have outcomes belonging to more than one decision outcome class (i.e., joint decision outcomes). Furthermore, when applied to domains with a single decision outcome, these algorithms become less efficient as the number of the pre-defined outcome classes increases. The objective of this dissertation is to modify previous decision-tree induction techniques in order to apply them to applications with joint decision outcomes. We propose a new decision-tree induction approach called the Multi-Decision-Tree Induction (MDTI) approach. Data was collected for a patient image retrieval application where more than one prior radiological examination would be retrieved based on characteristics of the current examination and patient status. We present empirical comparisons of the MDTI approach with the Backpropagation network algorithm and the traditional knowledge-engineer-driven knowledge acquisition approach, using the same set of cases. These comparisons are made in terms of recall rate, precision rate, average number of prior examinations suggested, and understandability of the acquired knowledge. The results show that the MDTI approach outperforms the Backpropagation network algorithms and is comparable to the traditional approach in all performance measures considered, while requiring much less learning time than either approach. To gain analytical and empirical insights into MDTI, we have compared this approach with the two best known symbolic learning algorithms (i.e., ID3 and AQ) using data domains with a single decision outcome. It has been found analytically that rules generated by the MDTI approach are more general and supported by more instances in the training set. Four empirical experiments have supported the findings.