How to deal with categorical variables in this method about gain HOT 8 CLOSED

I've also tried the same thing with categorical variables and the model seems to suffer from mode collapse (within a categorical variable). Seeking some guidance on this.

from gain.

HI!

Thank you for providing the code to your paper.
I've realized that in the implementation, the loss function used for the reconstruction loss only considers continuous variables in contrast with the paper that considers both cases.
Do you happen to have the code for the categorical variables too?

from gain.

Don't have the code but these are the tips from the author.

On top of one-hot encoding, to make categorical variable work:
(1) Activation function: use softmax activation for each spanned categorical vector (instead of sigmoid activation)
(2) Change the MSE loss to the cross-entropy loss for the categorical variable.

from gain.

Thanks, HaoXiao2018.
This is exactly what I did for the categorical variables.
As can be seen in the paper, I used cross-entropy for categorical variable and mse for continuous variable.
But more important thing is that we need to use separate output layer (activation function with softmax).

from gain.

So, is it possible to have both categorical and continuous features in a dataset?

from gain.

Yes. It would be better if you modify the codes based on the comments from HaoXiao2018 above and apply GAIN for your mixed-type data.

from gain.

If I have categorical data, does that mean I also have to change the dimensions of all the network layers, etc as well?

from gain.

Network parameters should be updated for different datasets.
Even for the datasets which only have continuous variables, those hyper-parameters should be optimized to maximize the performance.

from gain.