Deep causal trajectory modeling of public opinion evolution from cross-media information streams

Aiming at the shortcomings of traditional research methods in dealing with multi-source heterogeneous data and capturing the dynamic characteristics of concept evolution, a three-layer progressive framework of multimodal alignment, spatio-temporal modeling and causal inference is proposed. By constructing a meta-learning framework based on contrastive learning, the problems of modal heterogeneity and semantic fragmentation of cross-media data are solved, and the alignment and unified representation of different modal features are realized. Furthermore, the ST-Graphformer model is designed, which combines the graph neural network (GNN) and the Transformer architecture to effectively capture the temporal and spatial correlation characteristics in the process of concept evolution. At the same time, causal inference technology and attention mechanism are introduced to enhance the interpretability of the model and reveal the driving factors and key paths behind the evolution of ideas. The experimental results show that the proposed ST-Graphformer model performs well in emotion classification and trajectory prediction tasks. Compared with the traditional model and baseline model, the F1 score of emotion classification is increased by 11.8%, the DTW distance of trajectory prediction is decreased by 37.6%, and the recall rate of cross-modal retrieval is 0.63 at R@10. The ablation experiment verifies the contribution of each module to the model performance, and the case study further demonstrates the superior performance of the model in tracking and analyzing the evolution track of ideas and quantifying the causal effect of intervention variables.