Rule extraction from neural networks via decision tree induction

Rule extraction from neural networks is the task for obtaining comprehensible descriptions that approximate the predictive behavior of neural networks. Rule-extraction algorithms are used for both interpreting neural networks and mining the relationship between input and output variables in data. This paper describes a new rule extraction algorithm that extracts rules that contain both continuous (real-valued) and discrete literals. This algorithm decomposes a neural network using decision trees and obtains production rules by merging the rules extracted from each tree. Results tested on the databases in UCI repository are presented.