Understanding Neural Networks via Rule Extraction

Although backpropagation neural networks generally predict better than decision trees do for pattern classi cation problems they are of ten regarded as black boxes i e their predic tions are not as interpretable as those of deci sion trees This paper argues that this is be cause there has been no proper technique that enables us to do so With an algorithm that can extract rules by drawing parallels with those of decision trees we show that the predic tions of a network can be explained via rules ex tracted from it thereby the network can be un derstood Experiments demonstrate that rules extracted from neural networks are compara ble with those of decision trees in terms of pre dictive accuracy number of rules and average number of conditions for a rule they preserve high predictive accuracy of original networks