A named entity topic model for news popularity prediction

Yang Yang, Xiaoling Lu, Feifei WangOpenReview 2024 · OpenReview

TL;DR

This paper introduces a novel "Named Entity Topic Model" (NETM) to predict the popularity of news articles before they are published. By modeling how named entities (like people, places, organizations) contribute to popularity across different topics, NETM significantly outperforms existing methods, offering a more interpretable way to understand what drives news engagement.

What Problem Does This Paper Address?

Predicting how popular a piece of online content will be is a crucial but difficult task, especially for time-sensitive content like news articles. The challenge is amplified when trying to predict popularity before an article is even published (the "cold start" problem), as there's no early engagement data (like views or shares) to learn from. Traditional methods often rely on content features, but understanding the interplay between who or what is mentioned (named entities) and what the article is about (topics) is key to accurately forecasting popularity. Existing approaches haven't fully explored this interaction, leaving a gap in our ability to predict and explain news engagement.

Methods

The core innovation is the Named Entity Topic Model (NETM). NETM builds on standard topic modeling techniques but introduces a new layer: it models the "popularity gain" of each named entity (NE) across all topics. The popularity of an article is then seen as the sum of popularity gains from its NEs, weighted by their contribution to the article's topics.

Here's a breakdown of the NETM's generative process:

  1. Topic-Word Distributions: Standard topic modeling where each topic is defined by a set of words.
  2. NE-Topic Distributions: Each NE is assigned a probability distribution over topics, indicating its typical thematic associations.
  3. Article Topic Distributions: Each article has its own distribution of topics.
  4. Word Generation: Words in an article are generated by first picking a topic and then picking a word from that topic's distribution.
  5. Popularity Gain ($\lambda_d$): This is the crucial new part. It's calculated as a sum over all topics and all NEs present in the document. It represents the aggregated "popularity potential" derived from the interaction of NEs and topics.
  6. Observed Popularity ($Y_d$): The actual popularity (e.g., view count) is modeled as a Poisson distribution based on this calculated popularity gain.

To handle the vast number of NEs, the authors first cluster them into groups (e.g., "National Leaders," "Entrepreneurs") and use these clusters in the model, significantly reducing sparsity and computational burden.

Datasets

The study uses two large, real-world Chinese news datasets:

  • NetEase News: Collected from June 2016 to September 2018, containing 357,921 articles. Popularity indicators include view counts and viewer comments.
  • Tencent News: Collected from June 2018 to March 2019, with 87,083 articles. Popularity indicator is viewer comments.

These datasets are preprocessed by removing digits, punctuation, English words, performing word segmentation, removing stopwords, and then extracting NEs (people, places, organizations) using specialized toolkits.

Key Insights and Ideas

  • NEs and Topics are Intertwined Drivers of Popularity: The paper strongly suggests that it's not just what an article is about, but who or what is mentioned within those topics, that significantly influences its popularity. For instance, an article about "finance" might gain different traction depending on whether it features "entrepreneurs" or "government leaders."
  • Clustering NEs is Crucial for Scalability: Dealing with millions of individual NEs would lead to severe data sparsity. Clustering them into meaningful groups (like "National Leaders") makes the model tractable and allows for generalization.
  • Interpretable Popularity: The learned "NE topic probability matrix" provides insights into why certain articles become popular. It highlights which NE-topic combinations are associated with higher engagement.
  • "Cold Start" Prediction is Feasible: By focusing on intrinsic content features (NEs and topics) rather than early engagement metrics, NETM demonstrates effective prediction even before an article is published.

Key Results

  • Superior Prediction Accuracy: NETM consistently outperformed several baseline and state-of-the-art methods (including SVM-NE, S-LDA, LSTM, FC, and TPAP) across all tested datasets and popularity indicators (NetEase views, NetEase comments, Tencent comments) in terms of RMSE and MAPE. For example, NETM achieved significantly lower RMSE values compared to the next best methods.
  • Strong Performance on Well-Received Articles: When predicting the popularity of highly popular articles (top quantiles), NETM also showed the highest accuracy percentages.
  • Timeline Prediction: NETM maintained its superior performance even when predicting popularity over time, demonstrating robustness.
  • Category-Specific Performance: The model showed strong performance across different news categories, with categories having more NEs (like "Focus News" and "Entertainment") often yielding better prediction results, reinforcing the importance of NEs.

Limitations and Future Work

  • Computational Cost: NETM is computationally more expensive than simpler models like LDA, especially with a large number of NE clusters and documents. The authors suggest developing parallel algorithms for future work.
  • Mmax Sensitivity: The model's performance is influenced by the choice of the upper popularity limit ($M_{max}$). While they found an optimal range, careful tuning is required.
  • Future Directions:
    • Incorporating more textual features like sentiment.
  • Jointly modeling NE extraction and NETM.
  • Developing parallel computing algorithms for efficiency.
  • Exploring a more generalized "NE topic" matrix that can account for NEs that decrease popularity (currently, the model focuses on positive contributions).

Summary updated

Abstract

Predicting the popularity of web content is widely regarded as an important but challenging task. Online news articles are typical examples of this. In particular, owing to their time-sensitive nature, it is preferable to predict the popularity of news articles before their publication. To achieve this, this study proposes a named entity topic model (NETM) to extract the textual factors that can drive popularity growth. Here, each named entity is assumed to have a popularity-gain distribution over all semantic topics. The popularity of a news article is considered as the accumulation of popularity gains generated by its named entities (NEs) over all the topics. By learning the popularity-gain matrix for each named entity, the popularity of any news article can be predicted. Experiments on two collections of news articles demonstrate that the proposed NETM can outperform existing models in terms of accuracy. Additionally, the popularity-gain matrix learned by the NETM can be used to effectively explain the popularity of specific news articles.
BibTeX
@misc{yang2024named,
  author = {Yang Yang and Xiaoling LU and Feifei Wang},
  title = {{A named entity topic model for news popularity prediction}},
  year = {2024},
  url = {https://openreview.net/forum?id=T1fmiTw8n3},
  howpublished = {OpenReview},
}
Figure