A novel evolutionary deep reinforcement learning algorithm for the influence maximization problem in multilayer social networks

How to identify a set of influential individuals that can ensure the most information diffusion in multilayer social networks remains a fundamental yet underexplored issue of the influence maximization problem. Existing solutions mostly simplify or even neglect the heterogeneous characteristics of individuals from different layers, and the inter-layer propagation dynamics of the information spreading in the multilayer social networks. To address such challenges, a cross-layer independent cascade model is proposed to capture the inter-layer information cascading effect. Furthermore, this paper proposes a differential evolution-aided deep reinforcement learning (DEDRL) algorithm to identify the optimal seed set for the influence maximization in multilayer networks. More specifically, a multilayer network embedding mechanism is conceived to learn node embeddings of multilayer networks and the differential evolution is integrated with deep reinforcement learning to evolve a population composed of deep Q network weight parameters. Experimental evaluations conducted on both synthetic and real-world multilayer networks demonstrate the effectiveness of the proposed DEDRL and show an average performance improvement of 3.8% compared to the state-of-the-art algorithms.