News Recommendation in Forum-Based Social Media

Jia Wang, Qing Li, Yuanzhu Peter Chen, Jiafen Liu, Chen Zhang, Zhangxi LinAAAI 2010 · DOI

Abstract

Traditional news recommendation systems are often static, focusing primarily on the original article content. In the context of forum-based social media, discussions frequently evolve beyond the initial topic. This paper introduces an adaptive framework that treats news threads as dynamic conversations, incorporating user comments and structural discussion dynamics to recommend articles that align with the community's evolving interests while reducing redundancy.

Introduction

In modern social media platforms like Reddit or Digg, a news article is often just the starting point for a broader conversation. A post about a technological advancement might pivot into a debate on ethics or privacy. Standard recommendation systems typically fail in these environments due to two main flaws: topic shifting—ignoring the divergence in comments—and result redundancy—recommending near-identical articles that offer no new context. To address these issues, this research proposes a framework that acts as a "discussion moderator," adapting its recommendations to the current state of the thread.

The Adaptive Recommendation Framework

The proposed system moves beyond simple text analysis by modeling the discussion as a multi-relation graph. This approach allows the recommender to follow the "gist" of the conversation as it shifts away from the original headline.

The proposed adaptive framework for news recommendation, illustrating the flow from news and comments to relevant article retrieval.

The framework consists of several key components:

  • Multi-Relation Graph Modeling: The discussion is represented as a graph where nodes are the original post and comments. Three relationship types are considered: content similarity, reply-to hierarchy, and quotation relations.
  • Comment Ranking: A modified PageRank algorithm identifies "important" comments based on engagement (replies and quotes).
  • Adaptive Topic Profiling: The system builds a weighted term vector that balances the original author's intent with the readers' evolving interests.
  • Relationship Explanation: To combat redundancy, recommended articles are categorized as Generalizations, Specializations, or Duplications.

Experimental Evaluation

The framework was evaluated using a synthetic dataset derived from Digg (20 news postings and their comment threads) and a recommendation corpus of 16,718 Reuters articles. Human judges evaluated the top 200 results for each topic to establish ground truth for relevance and relationship types (generalization vs. specialization).

Results and Key Insights

The adaptive model significantly outperformed traditional baselines like Okapi and standard Language Modeling (LM).

Performance comparison of the proposed method using news only (RUN 1), comments only (RUN 2), and both (RUN 3).

Key findings include:

  • Superior Precision and Novelty: The combined model (RUN 3) achieved a P@10 of 0.94 and a Novelty (D@10) of 0.9, demonstrating its ability to find both relevant and non-redundant content.
  • The Power of Structural Logic: Structural data (who replied to whom) proved to be a more reliable indicator of a comment's importance than text similarity. Surprisingly, incorporating text similarity into the graph model sometimes introduced noise that decreased overall effectiveness.
  • Synthesis is Key: The ablation study confirmed that using both news content and comments yields the best results, effectively merging the author's perspective with the community's active interests.

Conclusion

This research highlights the necessity of treating social media news threads as living entities. By leveraging the structural dynamics of user discussions, recommendation engines can provide more contextually relevant and diverse content. Future work will focus on scaling the system for real-time applications and incorporating user reputation to further refine comment weighting.

Summary updated

BibTeX
@inproceedings{wang2010news,
  author = {Jia Wang and Qing Li and Yuanzhu Peter Chen and Jiafen Liu and Chen Zhang and Zhangxi Lin},
  title = {{News Recommendation in Forum-Based Social Media}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7502},
}
Figure