Abstract: The goal of influence maximization is to find a group of seed users who can increase the spread of information in social networks. Most current influence maximization methods focus on maximizing influence spread, often assuming that all social relationships are trustworthy. However, in reality, social networks are sensitive to trust. Users may trust, distrust, or even reject information depending on how trustworthy they see its source. Influence maximization techniques that neglect trust and distrust relationships risk propagating untrustworthy influence, which can be detrimental to overall performance. Therefore, this paper proposes a Trust-Aware Deep Reinforcement Learning (TA-DRL) framework for influence maximization in signed social networks. The social network is represented as a directed and weighted trust graph, where edges encode trust relationships. A trust-aware Graph Neural Network (GNN) is employed to learn node representations via trust-weighted message passing. The learned node representations serve as inputs to a deep reinforcement learning agent that sequentially selects seed nodes under a fixed budget constraint. The influence maximization problem is formulated as a Markov Decision Process (MDP), and a trust-aware reward function is designed to maximize trustworthy influence while minimizing negative influence. Extensive experiments are conducted on four real-world signed social network datasets: Epinions, Slashdot Zoo, Wiki-RfA, and Bitcoin-Alpha. The experimental results confirm that the proposed TA-DRL framework outperforms existing methods in terms of influence maximization, trust-aware influence maximization, and negative influence suppression. Runtime and scalability analyses further verify the effectiveness of the framework in achieving a favorable trade-off between performance and computational efficiency.

Key words: influence maximization, TA-DRL, GNN, deep reinforcement learning