From Text to Impact: A Deep Multimodal Approach for Predicting and Explaining Scholarly Influence Dynamics

The dynamics of scholarly influence are shaped by both semantic innovations in text content and structural patterns in citation networks. However, existing literature metrology methods often analyze these modalities in isolation, neglecting their temporal interplay and failing to explain prediction rationales. To address these limitations, this paper proposes a deep multimodal framework for predicting and explaining scholarly influence dynamics through joint modeling of text semantics and citation networks. Our approach integrates three key innovations: (1) a Temporal Graph Attention Network (TGAT) that captures evolving citation relationships and knowledge diffusion paths, (2) a BERT-Heterogeneous Graph Transformer (HGT) hybrid architecture that aligns semantic embeddings with topological node representations via cross-modal attention, and (3) a SHAP-based attribution mechanism that quantifies the marginal contribution of textual claims, citation contexts, and structural positions to influence propagation. Evaluated on a large-scale academic graph dataset spanning 500,000 computer science papers (2010–2023), the model achieves state-of-the-art performance in predicting future high-impact works (F1-score: 0.79) while revealing actionable insights: temporal citation bursts are found to amplify semantic novelty by 37%, and cross-disciplinary references contribute $2.4 \times$ higher influence than intra-disciplinary ones.