Reinforcement Learning–Enhanced Influence Maximisation via Credibility‐Based Clustering and Metaheuristic Optimisation

ABSTRACT

Influence maximisation (IM) seeks to identify a limited set of seed nodes that maximises information diffusion in social networks. Traditional centrality‐based and greedy approaches often rely on static structural assumptions and overlook community heterogeneity, behavioural credibility and adaptive optimisation. To address these limitations, this study integrates credibility‐based community detection, metaheuristic optimisation, and reinforcement learning‐driven parameter adaptation in a three‐phase framework. Evaluated on four datasets (Caida, Cond‐Mat, Douban and route views), the framework achieves 29%–68% influence spread and 11%–25% runtime reduction compared to degree centrality baselines. Ablation analysis confirms the incremental contribution of each component, demonstrating improvements from the base model (20% Caida, 3% route views) to the full model (29% Caida, 53% route views). First, type‐2 fuzzy clustering combines structural and behavioural features to identify cooperative communities and prioritise credible nodes. Second, the Ali Baba and the 40 Thieves (AFT) metaheuristic selects diverse and high‐impact seed nodes using a multi‐criteria fitness function that balances preferential attachment, neighbourhood overlap reduction and structural similarity. Third, a reward‐adaptive mechanism dynamically adjusts the fitness weights ( α , β and γ ) based on diffusion feedback, allowing the framework to adapt to varying network conditions within the tested regimes (subcritical, near‐critical and supercritical). The framework is evaluated on four real‐world datasets under a weighted independent cascade model across subcritical, near‐critical and supercritical regimes. Compared to the IMBC baseline, the framework achieves higher influence spread across all regimes, with ablation analysis quantifying each component's contribution: base model (3%–20% spread) → full dodel (29%–68% spread), representing 1.45× to 17.67× relative improvement.