Cellular networks are overloaded due to the mobile traffic surge, and mobile social network (MSNets) carrying information flow can help reduce cellular traffic load. If geographically-nearby users directly adopt WiFi or Bluetooth technology (i.e., leveraging proximity-based communication) for information spreading in MSNets, a portion of mobile traffic can be offloaded from cellular networks. For many delay-tolerant applications, it is beneficial for traffic offloading to pick some seed users as information sources, which help further spread the information to others in an epidemic-like manner using proximity-based communication. In this paper, we develop a theoretical framework to study the issue of choosing only k seed users so as to maximize the mobile traffic offloaded from cellular networks via proximity-based communication. We introduce a gossip-style social cascade (GSC) model to model the information diffusion process, which captures the epidemic-like nature of proximity-based communication and characterizes users' social participation as well. For static networks as a special-case study and mobile networks, we establish an equivalent view and a temporal mapping of the information diffusion process, respectively, leveraging virtual coupon collectors. We further prove the submodularity in the information diffusion and propose a greedy algorithm to choose the seed users for proximity-based traffic offloading, yielding a solution within about 63% of the optimal value to the traffic offloading maximization (TOM) problem. Experiments are carried out to study the offloading performance of our approach, illustrating that proximity-based communication can offload cellular traffic by over 60% with a small number of seed users and the greedy algorithm significantly outperforms the heuristic and random algorithms.