Solving the Influence Maximization Problem Via a Deep- Reinforcement- Learning-Guided Evolutionary Approach

The influence maximization (IM) problem aims to identify a set of seeds with the maximal influential ability within the network, where numerous applications can be detected such as market marketing and idea dissemination. In the early stages, the 1M problem has primarily been addressed using heuristic-based and population-based algorithms. Due to the NP-hard nature of this problem, these approaches are unable to meet the demands of large-scale networks. Previous work incorporates deep learning into this problem and algorithms have been developed based on Deep Q-Network (DQN). However, the performance in handling the seed determination task tends to be inferior to heuristic algorithms. For achieving competitive results, this paper proposes a seed selection method based on the reinforcement learning to be embedded in an evolutionary diagram, namely RLEA-IM, to tackle the IM problem. Through experiments conducted on both synthetic and real-world networks, RLEA-IM demonstrates a superior performance over existing DQN-based algorithms. It is capable of generating seed sets with better influential ability.