Extracting Knowledge from Artificial Neural Networks: An Empirical Comparison of Trepan and Symbolic Learning Algorithms

Extracting meaningful knowledge from Artificial Neural Networks (ANNs) is a current issue since, for several applications, the ability to explain the decisions taken by ANNs is even more important than their classification performance. Although some techniques have been suggested to solve this problem, a large number of these techniques can only be applied to some specific ANNs models. This paper proposes and investigates the use of symbolic learning algorithms, such as C.45, C4.5rules [13], and CN2 [4], to extract meaningful symbolic representations from trained ANNs. The main difference of this approach with other techniques previously proposed is that it can be applied to any supervised ANN model. The approach proposed is in some way similar to the one used by the Trepan algorithm [5], which extracts a symbolic representation, expressed as a decision tree, from a trained ANN. Experimental results are presented and discussed in order to compare the knowledge extracted from several ANNs using the proposed approach and the Trepan approach. Results are compared regarding two aspects: fidelity and comprehensibility. The results obtained show that our approach, using C4.5, C4.5rules and CN2 as symbolic learning algorithms, produces in general better comprehensible symbolic representation than Trepan for the trained ANNs considered in the experiments.