A deep reinforcement learning framework for influence maximization problem on large-scale social networks

Recently, the Influence Maximization (IM) problem has gained significant prominence in both industrial applications and academic research. The dissemination of information on social influence maximization has attracted significant attention from both industries and academia. Since the exponential expansion of network data and the increasing intricacy of application scenarios, traditional approximation algorithms face substantial challenges, including insufficient theoretical guarantees, suboptimal empirical performance, and prohibitive computational costs. Although several deep learning-based algorithms have emerged recently, they often exhibit limited generalizability when applied to heterogeneous large-scale networks. To tackle this issue, in this paper we introduce a Deep Reinforcement Learning framework, MaDGNN, to solve the Influence Maximization problem on large-scale social networks via graph neural network and reinforcement learning. Hence, our algorithm can be applied to different types of networks. Extensive empirical experiments in real-world scenarios demonstrate that our model surpass baseline methods by a significant margin across diverse datasets. Furthermore, the proposed algorithm trained on smaller graphs exhibits excellent scalability to larger graphs.