The Sensitivity analysis for Tabular solution provides the sensitivity of any output feature to the state of any node in the network, including the nodes of prime interest: input features.

This functionality is based on the input perturbation algorithm, the general idea of which is:

• Add a set of perturbations of different magnitudes to the node under consideration

• Calculate the corresponding variation of the output

• Repeat the process for each node independently

• Deduce the common rank of importance of nodes

The Sensitivity analysis give an insider look at the model’s logic and can sometimes reveal that some input features are more important than others. This can lead to changes in the choice of inputs features. If one of input features has a little impact on the outputs, it can be excluded from training. The new training with less input features generally leads to a smaller model without significant loss of accuracy.

The NeurEco Sensitivity analysis is performed in Network sensitivity sections of GUI and proposes two modes: