Symbolic Representation of Neural Networks

Although backpropagation neural networks generally predict better than decision trees do for pattern classiication problems, they are often regarded as black boxes, i.e., their predictions cannot be explained as those of decision trees. In many applications, more often than not, explicit knowledge is needed by human experts. This work drives a symbolic representation for neural networks to make explicit each prediction of a neural network. An algorithm is proposed and implemented to extract symbolic rules from neural networks. Explicitness of the extracted rules is supported by comparing the symbolic rules generated by decision trees methods. Empirical study demonstrates that the proposed algorithm generates high quality rules from neural networks comparable with those of decision trees in terms of predictive accuracy, number of rules and average number of conditions for a rule. The symbolic rules from nerual networks preserve high predictive accuracy of original networks. An early and shorter version of this paper has been accepted for presentation at IJCAI'95.