A Machine Learning-Based Prediction Model for Network Propagation

To address the limitations of traditional methods in predicting network propagation dynamics—namely low computational efficiency, insufficient predictive capability, and strong dependence on fixed time steps—this paper proposes a prediction model that integrates static topological features with a time-step adaptation mechanism. The model constructs a predictive framework by combining eight-dimensional static topological features of nodes with the time step, enabling accurate prediction of the infection rate at any given time when an arbitrary node serves as the infection source. Extensive validation is conducted on BA scale-free networks, WS small-world networks, and real-world networks. Experimental results demonstrate that the proposed model can accurately reproduce the infection process. For full time-step predictions of infection size, the LightGBM algorithm achieves a coefficient of determination (R2) consistently above 0.97 when compared to ground-truth data. Furthermore, this paper introduces NMA to characterize the model’s prediction accuracy for any node at any time step, with an average accuracy around 80%. The introduced single-node RMSE and MAE time-step heatmaps show maximum values below 0.3. At the application level, the model can predict future outcomes based on node topological structures.