Low number of inputs compared to outputs (per row) in neural network

Question

I have data that one row has input of size 1x2 (two values) and output is matrix of size 15x3 (fifteen rows with three values) like that

            |-----------------|
            |y01,2 y01,2 y01,3|
            |y02,1 y02,2 y02,3|
            |.                |
            |.                |
            |.                |
            |.                |
            |.                |
-------     |.                |
|x1,x2|     |.                |
-------     |.                |
            |.                |
            |.                |
            |.                |
            |.                |
            |y15,1 y15,2 y15,3|
            |-----------------|

This is one row and I have around 3000 inputs like that. Most values are real numbers except one output column that is binary (So output y$_{n,3}$ values are 0 or 1).

1. How should I preprocess this kind of data?

2. Is it of any sense to train two models, one for real numbers only and the second one for binary output?

3. What neural network architecture would work best for this task?

Answer 1

1.) Perform standard preprocessing for you input data, i.e. remove mean and scale to unit variance.

2.) No, you should train everything together. Although it is at the moment not clear, there can be relevant relationships between your outputs.

3.) You can actually use any architecture. Simply start with Dense-NN. More important is how you arange your outputs. You should have different activation and loss functions for your different outputs. For your binary outputs use (binary)-Crossentropy and for your real values use Mean Squared Error (or sth. similar). Your final output consists of course of the sum of your single losses.