Frame-attentive dynamic link prediction: uncovering narrative resonance in social networks with large language model

Predicting the evolution of dynamic social networks is a fundamental challenge in computational social science. Traditional methods that rely on network structure or surface level textual features often fail to capture the latent ideological drivers governing user interactions. In this paper, we propose that the formation of future social ties is primarily driven by alignment with evolving narrative frames rather than simple textual similarity. To test this hypothesis, we introduce a novel framework for frame-attentive dynamic link prediction. First, we develop an LLM driven pipeline that integrates hierarchical community detection with bottom up abstractive summarization to induce structured semantic frames from unstructured user discourse. Second, we construct a dual layer heterogeneous graph that explicitly models both social structure (user-user interactions) and ideological alignment (user-frame connections). Finally, we propose a frame-attentive graph representation learning mechanism to generate user embeddings that are grounded in this socio-semantic context. Extensive experiments on two large scale geopolitics datasets from Weibo and X (Twitter) demonstrate the superiority of our approach. Our model significantly outperforms a wide range of strong baselines, achieving up to a 33.5% improvement in AUC-ROC and an 89% Precision@100. Ablation studies and temporal analyses further confirm that the LLM-induced frames provide a more robust and predictive signal for social dynamics than conventional features. This work provides strong empirical evidence for the narrative resonance hypothesis and offers a powerful new paradigm for modeling the evolution of online social networks.