Abstract
Predicting the growth of information cascades is a critical task in social media analysis. Traditional models often rely solely on network structure, ignoring the rich semantic content of the information being shared. We present DCGT (Deep Cascade Graph with Text), a framework that integrates network structure with textual content using a novel dynamic gating mechanism. By adaptively balancing structural influence and topical relevance, DCGT significantly outperforms state-of-the-art baselines in predicting cascade size increments across diverse datasets.
Introduction
Information cascades—the sequences of shares or citations that define how content spreads—are influenced by both the social network's topology and the content itself. Previous state-of-the-art models, such as DeepCas, successfully modeled cascades as paths in a graph but remained "blind" to the text. However, a user's likelihood to share a specific item depends heavily on their personal interests, which are reflected in their past textual history. The challenge lies in mathematically fusing rigid social graph structures with high-dimensional, messy human language.
Methodology
DCGT represents a cascade graph as a set of paths generated via random walks, which are then processed by Gated Recurrent Units (GRUs). The model's architecture is defined by three core innovations:
- Textual Attention: Instead of simple averaging, an attention mechanism queries a user's history to highlight past content most relevant to the current item.
- Dynamic Gating Mechanism: This serves as the model's decision-making core. For every node, the gate calculates a Structural Informativeness score (based on network position) and a Textual Informativeness score (based on topical match).
- Fusion Layer: The gate blends structural embeddings, textual embeddings, and a global fallback embedding to create a robust final representation of the node.
Experimental Setup
The model was evaluated on two large-scale datasets:
- Twitter: Retweet cascades from 2016, using past tweets as node text and the original tweet as cascade text.
- AMiner: A citation network where node text consists of an author's previous paper titles and cascade text is the title of the cited paper.
Results and Discussion
DCGT consistently achieved the lowest Mean Squared Error (MSE) across all tested time horizons. On the Twitter dataset, for a 1-day prediction window, DCGT achieved an MSE of 1.945, significantly outperforming the structure-only DeepCas (2.102) and traditional feature-based methods (2.894).
Key Insights
- The "Famous" vs. the "Expert": The gating mechanism successfully distinguishes between users who spread information due to central network positioning and those who spread it due to topical expertise.
- Mitigating Cold Starts: For "rare users" with limited cascade history, textual information acts as a vital signal, allowing the model to maintain accuracy where structural data is sparse.
- Contextual Fusion: Unlike simple concatenation, the dynamic gate allows the model to treat every user-item interaction as a unique event, applying weights based on the specific context of the share.
Conclusion and Future Work
DCGT demonstrates that joint modeling of text and networks provides a more holistic and accurate view of information diffusion. Future research may involve integrating more sophisticated NLP architectures like Transformers to improve document representation and explicitly modeling the temporal dynamics (exact timing) of shares to further refine prediction accuracy.
Summary updated