Zhishuang He, Minzhi Yuan, Zhipeng Li, Yue Guo

Hunan Mechanical & Electrical Polytechnic

Abstract：

Aiming at the problem that existing social network information propagation models struggle to balance temporal dynamics and network structure, this paper models social networks as weighted directed graphs and proposes an optimized model (D-TS-GAT-Transformer) integrating temporal analysis and deep learning. Firstly, the Directed Graph Attention Network (DiGAT) is used to extract the weighted structural features of "in-edge neighbors", and the interaction intensity is combined to quantify the unidirectional influence differences between nodes. Secondly, the Transformer encoder is employed to capture long-term temporal dependencies, making up for the fitting defects of traditional linear models. Finally, a cross-modal attention fusion module is designed to dynamically allocate the weights of directed structural features and temporal features by combining node in-degree/out-degree features, so as to adapt to the needs of the entire propagation process. Experimental results show that the proposed model significantly outperforms existing mainstream methods in prediction accuracy, and can provide technical support for social network public opinion monitoring.

Key Words：

social networks; temporal analysis; deep learning; directed graph attention network; transformer