Effectively extracting rules from trained neural networks based on the characteristics of the classification hypersurfaces

The lack of efficient heuristic information is the fundamental reason that causes the low effectiveness of currently used approaches for extracting symbolic rules from trained feedforward neural networks. In this paper, a new measurement method of the classification power of attributes on the basis of the characteristics of the classification hypersurfaces is proposed, which is suitable for both continuous attributes and discrete attributes. Based on this new measurement method, a new approach for rule extraction from trained neural networks and classification problems with continuous attributes is proposed. A new method to adjust the complexity of neural network models based on the output range adjustment is also presented. The performance of the new approach is demonstrated by several computing cases. The results of the experiments prove that the approach proposed can improve the validity of the extracted rules remarkably compared with other rule extracting approaches, especially for the complicated classification problems.