Abstract
Predicting which web events will go viral is notoriously difficult due to the complex relationship between text, images, and public interest. This paper introduces the Self-Mutual excitation graph Neural network (SMN), a model that treats keywords as nodes in a global graph to predict popularity through "excitation mechanics." By enforcing sparsity, the model not only outperforms state-of-the-art methods but also provides an interpretable look at exactly which keywords are driving an event's success.
Introduction: The Challenge of Cold-Start Popularity
In an era of information overload, content platforms rely on popularity scores to rank and distribute news. However, predicting the popularity of a newly emerging event is a "cold-start" problem. Existing approaches often fail for two primary reasons: a lack of interpretability and weak multi-modal correlation. Most deep learning models are black boxes; they might predict a score correctly but cannot explain why a specific event is trending. Furthermore, the relationship between a news headline and its accompanying image is often non-linear and semantically "weak."
Methodology: The SMN Framework
The proposed SMN framework consists of two main branches designed to capture the multi-faceted nature of web popularity.
The Text Branch
This is the core innovation of the paper. Keywords from all events are organized into a unified graph where edges are weighted using Point-wise Mutual Information (PMI) to capture semantic relevance across the entire dataset. The model decomposes popularity into three additive parts:
- Self-Excitation: The inherent "heat" of individual keywords.
- Mutual-Excitation: The synergistic effect of keywords co-occurring, modeled via an exponential linear layer.
- Base Excitation: The general popularity of the broader topic.
To identify the most influential words, the authors use a Straight Through Estimator (STE) to apply a sparsity mask. This forces the model to "pick" the most important keywords, making the output interpretable.
The Image Branch
The image branch uses CLIP (Contrastive Language-Image Pre-training) to extract visual features. These features are processed through a Multi-Layer Perceptron (MLP) to contribute a visual popularity component to the final score. The final prediction is the sum of the base, self-excitation, mutual-excitation, and image-based scores.
Experimental Setup and Datasets
The researchers evaluated their model on two primary datasets:
- Hot Events (New): A benchmark released with this paper containing 4,000 events from the Baidu Hotlist. It includes titles, descriptions, images, and search indices. Unlike older datasets (like Twitter-3 or Sina Weibo-1), it provides rich multi-modal data.
- HEP-PH: A dataset of academic paper citations used to test the model's generalizability in a different domain (scientific popularity).
Key Results and Insights
The SMN model (specifically the version using Graph Attention Networks, SMN-GAT) significantly outperformed standard Click-Through Rate (CTR) models like DeepFM and xDeepFM.
- Accuracy: On the Hot Events dataset, SMN-GAT achieved a Mean Square Error (MSE) of 0.09, compared to 1.02 for xDeepFM.
- Ranking: In terms of Mean Average Precision (mAP), SMN-GAT reached 0.75, a substantial lead over the 0.63 achieved by the best-performing baseline.
- Ablation Study: Removing the image modality or the mutual-excitation module led to a measurable drop in performance, proving that both visual context and word-interaction modeling are essential.
Limitations and Future Work
While the model successfully combines text and images, the interaction between these two modalities is currently additive. Future work will explore more sophisticated "cross-modal" fusion techniques where images and text can excite each other directly, and improve visual representations to better capture the "viral" quality of images.
Summary updated