Influence maximization (IM) has attracted the attention of researchers as one of the hot topics in the field of social network studies. The IM problem is posed as a combinatorial optimization task with the aim of identifying a subset of influential individuals who can maximize the expected influence spread. Over the years, various techniques have been proposed to maximize influence, including heuristic methods, greedy algorithms, and metaheuristic optimization. Recently, the IM problem has been addressed by deep learning-based methods such as graph neural networks with generally satisfactory performance. Despite numerous studies on the use of GNN in solving IM, there are still many obstacles to developing effective and efficient algorithms in terms of fast computation, ensuring accuracy, minimizing memory usage, being adaptable to different network dimensions, and of course increasing the influence spread that need to be overcome. In order to develop existing methods and fill the remaining gaps, this research presents a method for maximizing influence in social networks using Extended Graph Convolutional Network (GCN). The proposed method is presented in two phases. First, in the training phase, three topological features Degree, Betweenness, and K-shell are extracted from the network graph and, after being concatenated together, are used as input to the developed GCN model. Then, using the Independent Cascade propagation model, the influence label of each node is generated and the model is trained. In the testing phase of the proposed method, the trained Extended GCN model predicts the influence of graph nodes and adaptively selects a Seed at each stage. The process of selecting seeds using the control parameter and the IC model continues until the Seed set is finally obtained. The main innovation of the proposed method is the use of the head attention mechanism for developing the GCN and the adaptive process in selecting seeds. Such a mechanism in the IM problem allows both local information and the overall structure of the network to be better learned. Finally, such a model selects the best seeds step by step. Simulation results show that the proposed method increases the Influence Spread by 5.23% when using the eight-headed mechanism and by 3.2% when using the four-headed mechanism.
