Hypothesis-Driven Constructive Induction in AQ17-HCI: A Method and Experiments

The developed AQ17-HCI system consistently outperformed, in terms of the prediction accuracy on new examples, all systems that it was compared to, including the AQ15 rule learning system, GREEDY3 and GROVE decision-list learning systems, and REDWOOD and FRINGE decision-tree learning systems.

Abstract

The proposed method for constructive induction searches for concept descriptions in a representation space that is being iteratively improved. In each iteration, the system learns concept description from training examples projected into a newly constructed representation space, using an Aq algorithm-based inductive learning system (AQ15). The learned description is analyzed to determine desirable problem-oriented modifications of the representation space. These modifications include generating new attributes, removing redundant or insignificant ones, and/or agglomerating attribute values into larger units. New attributes are constructed by assigning names to groups of the best-performing characteristic rules for each decision class, and then are used to define the representation space for the next iteration. This iterative process repeats until the created hypotheses satisfy a stopping criterion. In several experiments on learning discrete functions, the developed AQ17-HCI system consistently outperformed, in terms of the prediction accuracy on new examples, all systems that it was compared to, including the AQ15 rule learning system, GREEDY3 and GROVE decision-list learning systems, and REDWOOD and FRINGE decision-tree learning systems. Although the proposed method was developed for the Aq-based rule learning system, it can potentially be adapted to any other inductive learning system. In this sense, it represents a universal new approach to constructive induction.