Categorical inputs and standardisation in Neural Networks

[pvareschi]

Re: Applied Analytics Using SAS Enterprise Miner

1. Are categorical inputs represented as dummy indicator variables in a Neural Network (Chapter 5 of course notes)? If so, are those indicators standardised based on the setting of property "Network -> Input Standardisation" (see page 5-8 of course notes)?

2. In what scenarios would it make sense to use the other options in property "Input Standardization" (i.e. Range and MidRange)? (see page 5-8 of course notes)

[gcjfernandez]

Unfortunately the answer is not simple. It depends on the complexity of the data, type of neural network structure and the type of activation function etc, My recommendation is do a simple experiment with the your data with different standardization methods and determine which method is appropriate. Build a standard process flow diagram with this comparison (comparing different standardization) in SAS EM and routinely run this as a preliminary step before developing your NN.