GTBIM: A GraphSage–transformer-based approach for influence maximization with dynamic layer weighting in multilayer networks

Influence maximization refers to recognizing influential spreaders within complex networks, which is essential for comprehending and regulating diffusion dynamics. Identifying these nodes in multilayer networks remains a challenging and relatively unexplored problem and in deep learning-based algorithms include challenges such as the need for Scalable and high-quality embedding generation, the structural incompatibility of graph-structured data with neural network architectures, the inability to dynamically calculate the influence power of a node throughout the multilayer network and weight the layers. To overcome these shortcomings, we introduce a deep learning-based algorithm called GTBIM, which aims to detect influential nodes in multilayer networks. The GTBIM algorithm proposes a scalable model consisting of graph neural network and transformer neural network architecture. In the first stage, a new and scalable approach based on random walk and degree is developed to generate high-quality initial node embeddings. Then a GraphSage model is used to incorporate neighborhood information and enrich initial embeddings, thereby ensuring the compatibility and proper preparation of graph-structured data for subsequent neural network architectures. After that, a transformer encoder using its attention mechanism dynamically estimates the influence power of each node in entire multilayer network and the weight of each layer. Finally, the nodes that obtained highest estimated influence values are chosen as the seed set. Experimental results on datasets reveal that the GTBIM algorithm has substantially outperformed existing algorithms in terms of efficiency and other introduced metrics. These results underscore the capability of deep learning architectures in effectively tackling the influence Maximization problem in large-scale and multilayer social networks.