Abstract
Predicting the size of message cascades is critical in various applications, such as online advertising and early detection of rumors. However, most existing deep learning approaches rely on cascade observation, which hinders accurate cascade prediction before message posting. Besides, these approaches overlook personalized retweeting behaviors that reflect users' inclination to retweeting specific types of information. In this study, we propose a universal cascade prediction framework, namely Cascade prediction regarding Multiple Stage (CasMS), that effectively predicts cascade popularity across message generation stage as well as short-term and long-term stages. Unlike previous methods, our approach not only captures users' personalized retweeting behaviors but also incorporates temporal cascade features. We perform the experiments in datasets collected ourselves as well as public datasets. The results show that our method significantly surpasses existing approaches in predicting the cascade during the message generation stage and different time periods in the cascade dynamics.
Introduction
Predicting whether a social media post will go viral—a phenomenon known as cascade popularity prediction—is a cornerstone of digital strategy and platform safety. While many AI models can forecast a post's eventual reach once it starts gaining traction, predicting success before a message is even posted, or during its earliest moments, remains a significant challenge.
Most current popularity prediction models suffer from "observation bias," requiring a "warm-up" period of several hours to make an accurate forecast. This limitation renders them ineffective for the message generation stage, where content optimization is most valuable. Furthermore, existing models often treat social networks as monolithic entities, overlooking personalized retweeting behaviors. Individual users have distinct preferences; one might frequently share sports news while another focuses on technology. Without accounting for these individual inclinations and the specific content of the message, models struggle to predict which messages will trigger a massive "burst" of activity versus a slow crawl.
Methodology: The CasMS Framework
To address these challenges, we introduce CasMS (Cascade prediction regarding Multiple Stage). As illustrated in Figure 1, the framework consists of four integrated components designed to capture the interplay between content, user behavior, and temporal dynamics.
- Modeling Personalized Retweeting Behaviors: The model uses BERT to extract semantic features from message text and Skip-gram graph embeddings to capture users' social positions. It then calculates a "personalized retweet probability" for every user in the network by aligning their interests with the message content.
- Learning Cascade Snapshot Features: Utilizing a Graph Convolutional Network (GCN), the model simulates the diffusion process. The GCN takes the personalized retweet probabilities and the current "activation state" (who has already retweeted) to predict how influence might propagate to neighbors. Crucially, the GCN is mathematically constrained to ensure "monotonicity," mimicking real-world social contagion where the probability of activation increases as more neighbors engage.
- Learning Multi-Resolution Temporal Features: To handle the timing of cascades, we employ a Temporal Convolutional Network (TCN). TCNs are superior to traditional LSTMs at capturing long-range dependencies in time-series data, allowing the model to retain older patterns that might influence future growth.
- Joint Prediction Module: These features are integrated into Multi-Layer Perceptrons (MLPs) to output the predicted increment in cascade size.
Experimental Evaluation
We evaluated CasMS on four real-world datasets: Twitter (2012), Weibo2016, APS (citation cascades), and a newly collected Weibo2022 dataset covering activity between late 2022 and mid-2023. This modern benchmark allows us to test the model against current social media dynamics.
Key Results and Burst Prediction
CasMS demonstrated a clear lead over 10 state-of-the-art baselines, including DeepCas and CasTformer. On the MSLE (Mean Squared Logarithmic Error) metric, CasMS showed an average improvement of over 10% across all datasets.
A standout capability of CasMS is its performance in the "Generation Stage"—predicting popularity before a message is posted. By leveraging the personalized behavior module, it significantly outperformed feature-based deep learning models that lack such granular user modeling.
Furthermore, CasMS excels at predicting "bursty" cascades. As shown in Figure 3(b), in a case study from the Weibo2022 dataset, CasMS was the only model to accurately predict a sharp surge in retweets occurring 24-32 hours after posting. Other models predicted a flat or smooth decline, failing to capture the latent potential for a secondary wave of engagement triggered by specific user interactions.
Conclusion
CasMS represents a significant step toward proactive popularity prediction. By shifting the focus from aggregate network statistics to personalized user-content interactions, the framework provides more accurate forecasts across all stages of a cascade. Future work will explore dynamic social graphs and multimodal content (images/video) to further refine prediction accuracy in the rapidly evolving social media landscape.
Summary updated