Quote:
Originally Posted by qkgia
Dear Prof.
I have a question related to neural network.
Can we train a neural network with inputs containing both discrete and continuous values? Does it affect on backpropagation algorithm? If it does, can we two separate network one for discrete inputs and one for continuous inputs?
Thank you.

We can use both continuous and discrete variables in the same way, assuming the discrete variable is numerical, i.e., it is a number albeit discrete, and it has a natural order (e.g., rating of a movie or number of hits on a web site, rather than just a code for events with no order).
For other types of discrete variables, there are still ways to code them for the network so that they can be used together with continuous variables. For instance, a variable that assumes 3 values that have no order can be coded as three binary variables, where only one of the variables is 1 and the others are 0 to designate one of the three values of the original variable.
In all of these cases, it often helps to normalize the numerical range of the variables (say to mean 0 and variance 1) to avoid some of them dominating backpropagation by their shee numerical strength.