Comparing methods to extract the knowledge from neural networks

Neural networks (NN) have been shown to be accurate classifiers in many domains. Unfortunately, the lack of NN's explanatory capability of knowledge learned has somewhat limited their application. A stream of research has therefore developed focusing on knowledge extraction from within neural networks. The literature, unfortunately, lacks consensus on how best to extract knowledge from help neural networks. Additionally, there is a lack of empirical studies that compare existing algorithms on relevant performance measures. Therefore, this study attempts to help fill this gap by comparing two different approaches to extracting IF-THEN rules from feedforward NN. The results show a significant difference in the performance of the two algorithms depending on the structure of the dataset utilized.