A influence prediction method in social networks with perturbation-constrained learning

The problem of influence prediction holds significant theoretical and practical importance in social network analysis, with wide applications in domains such as information dissemination, viral marketing, and public opinion monitoring. Traditional methods (such as degree centrality, betweenness centrality) rely on network topology for seed selection but suffer from inefficiency and incomplete evaluation due to their neglect of the collective diffusion dynamics of seed sets. Meanwhile, graph-based models trained on static networks (such as GNNs, GCNs) struggle to generalize to real-world dynamic scenarios. To address these issues, this paper proposes ADGATQL, a social network influence prediction method based on perturbation-invariant constraints. The approach first employs heterogeneous network pretraining to capture universal topological features, incorporates adversarial perturbations to generate diversified network variants, and integrates uncertainty-guided active learning to achieve efficient annotation. Finally, it dynamically optimizes the seed set through deep reinforcement learning, significantly reducing computational complexity while maximizing propagation coverage. Experiments on six real-world datasets demonstrate that the proposed method effectively resolves the challenges of insufficient adaptability and excessive computational costs faced by existing approaches in social networks.