An opinion evolution prediction model based on graph attention and ordinary differential equations

Accurately predicting the dynamic evolution of user opinions on social media is important for understanding opinion diffusion mechanisms, identifying patterns of collective behavior, and supporting public decision-making. Existing opinion dynamics models provide interpretable descriptions of opinion evolution but rely on idealized assumptions and may not adequately capture the structural complexity of real-world social networks. In contrast, purely data-driven methods have strong approximation capabilities but lack explicit modeling of underlying social interaction processes. To address these limitations, we propose GA-ODE, an opinion evolution prediction model that integrates a graph attention mechanism with neural ordinary differential equation (ODE) constraints. GA-ODE adaptively learns differentiated aggregation weights for neighboring nodes through a graph attention mechanism and employs a gated evolution unit to balance individual opinion persistence and neighborhood assimilation effects. We further introduce a structure-aware ODE consistency regularizer based on probabilistic bounded confidence and incorporate the adjacency matrix into the dynamical residual as a topology-aware mask. Experimental results show that the topology-aware mask provides limited benefits on clean graphs but improves prediction performance under structural perturbations and helps maintain dynamical consistency in degraded network conditions. Evaluations under non-idealized dynamical settings further show that GA-ODE maintains prediction performance in the presence of opinion repulsion effects and external event shocks. Experiments on multiple synthetic datasets and one real-world dataset validate the effectiveness and robustness of GA-ODE. However, the model assumes that social interactions remain static within the observation window and is sensitive to sparsely represented classes.