<p>Identifying influential nodes that can maximize information diffusion is a fundamental problem in social network analysis, commonly referred to as influence maximization. Diffusion model-based and traditional greedy and heuristic algorithms tend to be expensive to calculate, or poorly aware of the network structure, especially when dealing with large networks. To overcome these weaknesses, this paper will introduce a hybrid architecture that will combine Graph Neural Networks (GNNs) with the Ant Colony Optimization (ACO) towards influence maximization. The proposed approach works as follows, where the GNN learns influence representations on the node level, which jointly models both local and higher-order structural dependencies in the network, and ACO uses these learned representations to assemble a non-redundant and diverse seed set. The chosen seed sets are tested on the basis of the Independent Cascade diffusion model where the spread of influence is taken as the key performance measure. The experiments on the real-world social network data indicate that the proposed GNN-ACO framework is steadily more influential in spreading than the current baseline methods and that it has comparative computational efficiency at varying levels of seed set sizes. These findings demonstrate the advantage of optimizing influence maximization by using representation learning with a seed selection strategy based on optimization.</p>