Learning diffusion model-free and efficient influence function for influence maximization from information cascades

When considering the problem of influence maximization from information cascades, one essential component is influence estimation. Traditional approaches for influence estimation generally follow a two-stage framework, i.e., learn a hypothetical diffusion model from information cascades and then calculate the influence spread according to the learned diffusion model via Monte Carlo simulation or heuristic approximation. The effectiveness of these approaches heavily relies on the correctness of the diffusion model, suffering from the problem of model misspecification. Meanwhile, these approaches are inefficient when influence estimation is conducted via lots of Monte Carlo simulations. In this paper, without assuming a diffusion model a priori, we directly learn a monotone and submodular influence function from information cascades. Once the influence function is obtained, greedy algorithm is applied to efficiently solve influence maximization. Experimental results on both synthetic and real-world datasets show the effectiveness and efficiency of the learned influence function for both influence estimation and influence maximization tasks.