On the feasibility of utilizing correlations between user populations for traffic inference