Deep learning framework with Hadamard-based feature fusion for node influence power prediction

In this paper, an innovative architecture based on deep neural networks is presented. Initially, node and layer features are extracted as feature vectors. Each vector is then passed through a deep multilayer perceptron (MLP) network for enrichment. Using the Hadamard product, these vectors are multiplied element-wise to form a matrix. In the next step, to analyze feature interactions, this matrix is fed into a series of Transformer encoders arranged sequentially. Finally, an MLP network is used as a regression model to predict the influence power of the nodes. • A deep learning-based framework that leverages node and layer feature representations. • Integrating MLP networks and transformer encoders to offer a precise and computationally efficient approach. • Proposing a set of novel layer-specific features to enable effective layer weighting within the proposed framework.