tensorflow, multiple outputs w/ non-overlapping training data

Question

What is the correct method for training a tensorflow model with multiple outputs where the training examples of each output may not occur on the same record.

For example,