Influence Maximization with Bandits

Most work on influence maximization assumes network influence probabilities are given. The few papers that propose algorithms for learning these probabilities assume the availability of a batch of diffusion cascades and learn the probabilities offline. We tackle the real but difficult problems of (i)learning in influence probabilities and (ii) maximizing influence spread, when no cascades are available as input, by adopting a combinatorial multi-armed bandit (CMAB) paradigm. We formulate the above problems respectively as network exploration, i.e., minimizing the error in learned influence probabilities, and minimization of loss in spread from choosing suboptimal seed sets over the rounds of a CMAB game. We propose algorithms for both problems and establish bounds on their performance. Finally, we demonstrate the effectiveness and usefulness of the proposed algorithms via a comprehensive set of experiments over three real datasets.