Learning Influence Functions from Incomplete Observations