Multiple Agents Reinforcement Learning Based Influence Maximization in Social Network Services

Yiming Liu, Waichau Sze, Xiaofeng Gao, Guihai Chen

Abstract: Influence Maximization (IM), an NP combinatorial optimization problem, has been broadly studied in the past decades. Existing algorithms for IM are still limited by accuracy, scalability and generalization. Moreover, they solve the influence overlapping problem implicitly. This paper proposes Multiple Agents Influence Maximization (MAIM) scheme, a novel Machine Learning based method for IM problem. We focus on explicitly solving the influence overlapping hidden in IM. MAIM first generates a list of sorted nodes as seed candidates in a descending order of overall influence, and drops those with serious influence overlapping based on multiple reinforcement learning (RL) agents in different rounds. We make full use of the characteristics of RL agents: continuous interaction with the environment, quick decision on whether a node should be accepted or dropped and better generalization. We also propose Memory Separated Deep Q-Network to improve training efficiency. Experiments on eight real-world social networks validate the effectiveness and efficiency of our algorithm compared to state-of-the-art algorithms.

External IDs:dblp:conf/icsoc/LiuSGC21