Abstract
COVID-19 has affected the world economy and the daily life routine of almost everyone. It has been a hot topic on social media platforms such as Twitter, Facebook, etc. These social media platforms enable users to share information with other users who can reshare this information, thus causing this information to spread. Twitter's retweet functionality allows users to share the existing content with other users without altering the original content. Analysis of social media platforms can help in detecting emergencies during pandemics that lead to taking preventive measures. One such type of analysis is predicting the number of retweets for a given COVID-19 related tweet. Recently, CIKM organized a retweet prediction challenge for COVID-19 tweets focusing on using numeric features only. However, our hypothesis is, tweet text may play a vital role in an accurate retweet prediction. In this paper, we combine numeric and text features for COVID-19 related retweet predictions. For this purpose, we propose two CNN and RNN based models and evaluate the performance of these models on a publicly available TweetsCOV19 dataset using seven different evaluation metrics. Our evaluation results show that combining tweet text with numeric features improves the performance of retweet prediction significantly.
Introduction
During the COVID-19 pandemic, Twitter became a primary battleground for information and misinformation. Predicting which tweets will "go viral" via retweets is crucial for emergency management and tracking the spread of news. Most existing models—including those used in the prominent CIKM-20 Retweet Prediction Challenge—rely almost exclusively on numeric metadata: follower counts, friend counts, or timestamps. This research argues that such approaches ignore the most fundamental part of a tweet: the content itself. The authors demonstrate that the semantic meaning within tweet text provides vital context that numeric features alone cannot capture.
Methodology
The authors proposed two deep learning architectures to tackle retweet prediction as a regression task:
- CNN Regressor: This model uses One-Dimensional Convolution (Conv-1D) to extract local patterns from text, followed by k-max pooling (which picks the most "active" features) and folding (a technique to sum rows and reduce dimensionality).
- RNN Regressor: A simpler architecture using a Recurrent Neural Network layer with 32 hidden units and a tanh activation function, designed to capture sequential dependencies in the text.
The researchers tested these models in three configurations:
- Text-only: Using only the tweet content.
- Numeric-only: Using 12 features (followers, friends, favorites, mentions, sentiment scores, and split timestamp data).
- Combined: A dual-branch architecture where text and numeric features are processed separately, concatenated, and then passed through a final dense layer to predict the retweet count.
Datasets and Content Diversity
The study utilized the TweetsCOV19 dataset, a massive collection of over 8 million COVID-19 related tweets. Because many public datasets only provide tweet IDs, the authors "hydrated" the data to retrieve the actual text for approximately 6.9 million tweets. For the final experiments, they used a randomly sampled subset of 60,000 tweets.
As shown in the figure above, the text of COVID-19 tweets varies significantly, ranging from official health warnings and personal anecdotes from medical staff to misinformation regarding treatments and virus origins. This diversity underscores why text analysis is essential for understanding virality.
Key Insights
- The "Social Proof" vs. "Content" Balance: Numeric features (like follower count) generally performed better than text features when used in isolation. This suggests that who says something often matters more for retweets than what they say. However, the highest accuracy only appeared when both were combined.
- CNNs Outperform RNNs for Text Regression: In this task, the CNN model consistently beat the RNN. Retweetability is often driven by specific "trigger words" or phrases (local features) rather than the long-term narrative structure of a short message.
- Sentiment Nuance: While sentiment scores are often used as features, raw text captures nuances that a simple "positive/negative" score misses.
Results
The CNN Combined model was the clear winner across all evaluation metrics.
- Accuracy: The combined CNN model achieved an $R^2$ score of 0.7427, a significant jump from the 0.6466 achieved using numeric features alone.
- Error Reduction: The Mean Absolute Error (MAE) dropped from 30.81 (numeric-only) to 25.38 (combined) in the CNN model.
- The RNN Gap: The RNN performed poorly on text-only features ($R^2$ of 0.004), reinforcing that convolutional architectures are better suited for the short, punchy nature of tweet text.
Conclusion and Future Work
While the integration of text and numeric features significantly improves prediction, the authors identify multimodal content as a major remaining hurdle. Many viral tweets include videos or images that the current model cannot "see." Future research will focus on:
- Visual Features: Integrating image and video analysis.
- User History: Incorporating past posting behavior to model influence beyond raw follower counts.
- Graph Analysis: Examining the follower/following relationship graph to understand network-level information spread.
Summary updated