Decision Tree Extraction from Trained Neural Networks

Artificial Neural Networks (ANNs) have proved both a popular and powerful technique for pattern recognition tasks in a number of problem domains. However, the adoption of ANNs in many areas has been impeded, due to their inability to explain how they came to their conclusion, or show in a readily comprehendible form the knowledge they have obtained. This paper presents an algorithm that addresses these problems. The algorithm achieves this by extracting a Decision Tree, a graphical and easily understood symbolic representation of a decision process, from a trained ANN. The algorithm does not make assumptions about the ANN’s architecture or training algorithm; therefore, it can be applied to any type of ANN. The algorithm is empirically compared with Quinlan’s C4.5 (a common Decision Tree induction algorithm) using standard benchmark datasets. For most of the datasets used in the evaluation, the new algorithm is shown to extract Decision Trees that have a higher predictive accuracy than those induced using C4.5 directly.