Voting Consensus Incentive Mechanism Based on Deep Reinforcement Learning in Social Network Group Decision-Making

In practice, voting is the most common group decision-making (GDM) activity. The voting of each decision maker is usually driven by individual utility and influenced by his/her neighbors in social networks. The moderator aims to efficiently guide and incentivize the decision makers to reach a group solution with consensus. In the social network GDM (SNGDM), the individual preferences of decision makers and the influencing mechanisms of the social network are crucial to the achievement of voting consensus, but usually unobserved by the moderator. For the voting scenario with unobserved individual preferences and influencing mechanisms, this study presents a dynamic voting consensus incentive mechanism based on deep reinforcement learning (DRL) in SNGDM, which aims to maximize the number of decision makers within consensus under a limited budget. First, the voting consensus reaching process (CRP) is systematically modeled as a constrained Markov decision process (CMDP) from the perspective of the moderator, where the votes of decision makers are observed at each stage. Under an unknown decision environment (viz., unobserved individual preferences and influences in the social network), we propose a DRL-based consensus framework named constrained hybrid actor-critic (CHAC) to model the incentive behavior of the moderator with a limited budget. Then, a primal-dual deep deterministic policy gradient algorithm is developed to optimize the dynamic incentive mechanism of the moderator for the voting CRP. Finally, we conduct a thorough experimental evaluation on both synthetic and real social network datasets to quantify the performance of the proposed CHAC.