Learning Sequential Correlation for User Generated Textual Content Popularity Prediction

Wen Wang, Wei Zhang, Jun Wang, Junchi Yan, Hongyuan Zha

Abstract

Predicting the popularity of user-generated content (UGC) is a critical task for information prioritization on social media platforms. Traditional approaches often treat individual posts as independent events, ignoring the historical context of the author. This paper introduces the User Memory-augmented recurrent Attention Network (UMAN), a novel framework that models a user's posting history as a sequence. By leveraging external user memories to track evolving content preferences and popularity trends, UMAN captures both sequential content and popularity correlations. Evaluations on real-world datasets from Douban and Twitter show that UMAN significantly outperforms state-of-the-art baselines in predicting post engagement.

Introduction

In the age of information overload, platforms must effectively filter and prioritize content to maintain user engagement. Predicting "popularity"—measured by metrics such as retweets on Twitter or event RSVPs on Douban—is central to this effort.

Most existing research focuses on the features of a single post in isolation. However, this ignores two vital patterns inherent in human behavior:

  1. Sequential Content Correlation: Users often post about related topics within short timeframes (e.g., a series of posts about a tech conference).
  2. Sequential Popularity Correlation: A user’s influence or "clout" tends to evolve smoothly rather than randomly. Past engagement levels are strong predictors of future performance.

The challenge lies in designing a model that can "remember" these user-specific trends and apply them dynamically to new content.

Empirical Motivation: Quantifying the "Vibe"

To validate the hypothesis of sequential correlation, the authors analyzed how the relationship between two posts by the same user changes over time. As shown in the figure below, as the interval ($\Delta t$) between posts increases, content similarity (measured by cosine similarity of text embeddings) decreases, while the difference in popularity scores grows. This empirical evidence confirms that recent history is more indicative of current behavior and success than distant history, justifying a sequential modeling approach.

Content similarity and popularity difference vs. interval Δt.

The UMAN Architecture

The proposed User Memory-augmented recurrent Attention Network (UMAN) integrates several components to process sequential user data:

  • Text Representation: A Convolutional Neural Network (CNN) with multiple filter sizes (unigrams to trigrams) extracts phrase-level features from the post text.
  • External User Memory: A persistent vector ($v_n$) stores long-term user characteristics.
  • User-Aware Attention: The model uses the user's memory to weight the importance of specific words in a new post. For instance, a user with a history of technical posting will have an attention mechanism biased toward technical terminology.
  • Recurrent Modeling (LSTM): The $M$ most recent posts (text and popularity) are processed by an LSTM. This updates the user memory to reflect "short-term tendencies" or engagement streaks.
  • Popularity Prediction: The final popularity score is predicted by combining the updated user memory with the target post's representation through a fully connected layer.
Overview of the UMAN model architecture.

Experimental Evaluation

The model was tested on three datasets: D-Beijing and D-Shanghai (Douban Event data) and a large-scale Twitter dataset. Users were filtered to include those with at least 10 posts to ensure sufficient sequential context.

Key Results

  • Superior Performance: UMAN consistently achieved lower Mean Squared Error (MSE) than non-sequential baselines. For example, on D-Beijing, UMAN reached an MSE of 0.7197, significantly better than the 1.1949 achieved by the non-sequential HF-NMF model.
  • Feature Importance: The study found that incorporating both past text and past popularity scores provided the best results. Removing past popularity data led to a measurable performance drop, confirming that "success breeds success" in social media engagement.
  • Computational Efficiency: By using a CNN for initial text embedding rather than a purely recurrent approach for all text, the authors achieved a 5-10x speedup on Twitter data while maintaining comparable accuracy.

Conclusion and Future Work

UMAN demonstrates the power of treating user history as a continuous sequence rather than a collection of independent events. While the model shows strong gains, performance on short-form, noisy data like Twitter remains a challenge compared to more structured platforms like Douban. Future research could improve the model by explicitly accounting for the actual time elapsed between posts (rather than just their order) and incorporating explicit topic modeling to further refine content correlations.

Summary updated

Abstract

Popularity prediction of user generated textual content is critical for prioritizing information in the web, which alleviates heavy information overload for ordinary readers. Most previous studies model each content instance separately for prediction and thus overlook the sequential correlations between instances of a specific user. In this paper, we go deeper into this problem based on the two observations for each user, i.e., sequential content correlation and sequential popularity correlation. We propose a novel deep sequential model called User Memory-augmented recurrent Attention Network (UMAN). This model encodes the two correlations by updating external user memories which is further leveraged for target text representation learning and popularity prediction. The experimental results on several real-world datasets validate the benefits of considering these correlations and demonstrate UMAN achieves best performance among several strong competitors.
BibTeX
@inproceedings{wang2018learning,
  author = {Wen Wang and Wei Zhang and Jun Wang and Junchi Yan and Hongyuan Zha},
  title = {{Learning Sequential Correlation for User Generated Textual Content Popularity Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/225},
}
Figure