Generalized Additive Models from a Neural Network Perspective

Abstract

Recently, an interactive algorithm was proposed for the construction of generalized additive neural networks. Although the proposed method is sound, it has two drawbacks. It is subjective as it relies on the modeler to identify complex trends in partial residual plots and it can be very time consuming as multiple iterations of pruning and adding neurons to hidden layers of the neural network have to be done. In this article, an automatic algorithm is proposed that alleviates both drawbacks. Given a predictive modeling problem, the proposed strategy uses heuristic methods to identify optimal or near optimal generalized additive neural network topologies that are trained to compute the generalized additive model. The neural network approach is conceptually much simpler than many of the other approaches. It is also more accurate as heuristic methods are only used in identifying the appropriate neural network topologies and not in computing the generalized additive models.