How to determine the unique contributions of input-variables to the nonlinear regression function of a multilayer perceptron

• Shortcomings of previous publications on the topic are outlined (e.g., focus on linear relations).
• A simulation study is presented to investigate the case of non-monotonic relations.
• Results contradict conclusions of previous studies.

There are different methods for quantifying the relative contribution of input-variables to the nonlinear regression function provided by a Multilayer Perceptron. Unfortunately most of the systematic method comparisons available to date suffer from a set of characteristic shortcomings. This paper elaborates on these methodological shortcomings and presents a simulation study that demonstrates how to avoid them in future method comparisons. Results of the simulation study indicate that Garson's weight method is preferable to the connection weight method proposed by Olden et al., (2004) for each of the samples simulated.