Macroscopic diffusion prediction in social networks based on spatio-temporal and trend features

Abstract

Predicting the scale of information diffusion in social networks can sense the future propagation of information in advance, which plays a crucial role in controlling the diffusion of harmful information. We propose STTFP (Spatio-Temporal and Trend Features for Prediction), a deep learning framework that integrates temporal, spatial, and trend features to improve macroscopic diffusion prediction accuracy. This framework first utilizes graph attention networks to extract node interaction features from cascaded graphs. It captures node position features from diffusion sequences, and uses sparse matrix factorization to extract node features from social network graphs. Then it adopts bi-directional gated recurrent units and self-attention mechanisms to deeply mine spatio-temporal features. Additionally, we design an attention-based convolutional neural network to capture the short-term fluctuations in the information propagation process, while long short-term memory networks are used to uncover historical forwarding variation in diffusion scales. By fusing these features, the framework achieves incremental predictions of information diffusion. Experiments on three public datasets show that our method effectively enhances the accuracy of macroscopic diffusion predictions.