Symbolic rule extraction from the DIMLP neural network

Abstract

The interpretation of neural network responses as symbolic rules is actually a diffcult task. Our first approach consists in characte-rising the discriminant hyper-plane frontiers. More particularly, we point out that the shape of a discriminant frontier built by a standard multi-layer perceptron is related to an equation with two terms. The first one is linear, and the second is logarithmic. This equation is not suffcient to easily generate symbolic rules. So, we introduce the Discretized Interpretable Multi Layer Perceptron (DIMLP) model that is a more constrained multi-layer architecture. From this special network, rules are extracted in polynomial time and continuous attributes do not need to be binary transformed. We apply DIMLP to three applications of the public domain in which it gives better average predictive accuracy than C4.5 algorithm and discuss rule quality.