Here I have a multi-task network to detect if there is a talking user and where he is in a clip of audio, which has two outputs. One of them is a binary classification problem indicating if there is a user. Another output is a regression problem used to estimate the direction of the user. The input is a feature vector extracted from this audio clip.
If the classification output is 1, which means there is a user, then the output of regression has a meaningful value. So I want to set the target label data. The target label of classification can be easily set to 0 or 1. But I do not know how to set the label for regression for no-user cases. Is it valid to set it as a negative value (to differ it from the positive meaningful direction value)?
I know in a classification problem, we can set another new class to represent the no-exist case. How to do this in a regression model?