Data-driven prediction of information diffusion in online social networks using sparse dynamic mode decomposition

Predicting how information spreads in online social networks requires models that can handle both evolving network structures and dynamic user interactions. Existing popular approaches, either based on variants of epidemic models or learning from raw data, either oversimplify the network dynamics or demand high computational resources, and often result in limited interpretability. This paper proposes a data-driven prediction of information diffusion in dynamic online social networks, based on Sparse Dynamic Mode Decomposition (Sparse DMD). This model learns from structural and diffusion dynamics captured from temporal network snapshots by extracting dominant spatio-temporal modes that capture both local and global patterns of information spread. Experimental results from five real-world datasets (Higgs, Facebook Forum, Digg, Email-EU, Wiki-Talk) show that our approach consistently outperforms classical machine learning baselines on isolated prediction tasks (Level 1) by 15-30%. Compared to complex deep learning architectures on the joint co-evolution forecasting problem (Level 2), Sparse DMD achieves accuracy within 1.3%, while cutting computational costs by a factor of 7 to 30. The modal decomposition provides interpretable insights into dominant diffusion pathways and temporal trends. These results establish Sparse DMD as an efficient, interpretable, and scalable framework for understanding and forecasting information diffusion in dynamic social systems.