Generalized Analytic Rule Extraction for Feedforward Neural Networks

Abstract

This is paper suggests the "Input-Network-Training-Output-Extraction-Knowledge" framework to classify existing rule extraction algorithms for feedforward neural networks. Based on the suggested framework, we identify the major practices of existing algorithms as relying on the technique of generate and test, which leads to exponential complexity, relying on specialized network structure and training algorithms, which leads to limited applications and reliance on the interpretation of hidden nodes, which leads to proliferation of classification rules and their incomprehensibility. In order to generalize the applicability of rule extraction, we propose the rule extraction algorithm Generalized Analytic Rule Extraction (GLARE), and demonstrate its efficacy by comparing it with neural networks per se and the popular rule extraction program for decision trees, C4.5.