Data-Driven Theory Refinement Using KBDistAl

Abstract

Knowledge based artificial neural networks offer an attractive approach to extending or modifying incomplete knowledge bases or domain theories through a process of data-driven theory refinement. We present an efficient algorithm for data-driven knowledge discovery and theory refinement using DistAl, a novel (inter-pattern distance based, polynomial time) constructive neural network learning algorithm. The initial domain theory comprising of propositional rules is translated into a knowledge based network. The domain theory is modified using DistAl which adds new neurons to the existing network as needed to reduce classification errors associated with the incomplete domain theory on labeled training examples. The proposed algorithm is capable of handling patterns represented using binary, nominal, as well as numeric (real-valued) attributes. Results of experiments on several datasets for financial advisor and the human genome project indicate that the performance of the proposed algorithm compares quite favorably with other algorithms for connectionist theory refinement (including those that require substantially more computational resources) both in terms of generalization accuracy and network size.