I use scale normalization for the features of my neural network. As many of the values occur very infrequently, i.e. the minimum count within all samples is often also the count of the feature within many samples, I end up with a lot of zeros in my sample. E.g., out of 15 features 8 to 10 might be zero.
My question is: Is this is still ok, or should I add some smoothing parameter, in order to let the infrequent items also contribute?
