École Nationale Supérieure de
Techniques Avancées

Abstract

Neural networks are statistical models known for their great learning performances, but they are "black boxes": their decision making is not easily understandable by humans. On the contrary, the generalized additive model (GAM) is a classical statistical model that is very intelligible but has limited performances compared to neural networks. This work aims to study a new model developed: the neural additive model (NAM) which combines the power of neural networks with the clarity of GAMs. Many parameters were tested in order to understand their impact on the performance of this new model. Figures summarizing the results show the possible dependence between the accuracy score, the computation time and the parameters. This new model, although not yet optimized for its computation time, has comparable performances to other statistical models using decision trees.