Just being curious about adding noise to categorical variables. I suspect that adding gaussian noise over a one hot encoded variable wouldn't be enough. I was thinking about adding noise after embedding the categorical variables. After trying this, my results are not really improving. I suspect that the best solution would be to switch the one hot encoded variable randomly to another column. However it is unclear how to achieve this with tf keras.
Are there any other approach to adding noise to a categorical variable ? Regarding the last idea, is there any easy way to do this or do I have to build a custom layer ?