Strategies to choose nodes on a social network to maximize the total influence has been studied for decades. Studies have shown that the greedy algorithm is a competitive strategy and it has been proved to cover at least 63% of the optimal spread. Here we propose a learning-based framework for influence maximization aiming at outperforming the greedy algorithm in terms of both coverage and efficiency. The proposed reinforcement learning framework combining with a classification model not only alleviates the requirement of the labelled training data, but also allows the influence maximization strategy to be developed gradually and eventually outperforms a basic greedy approach.