PP-Rec: News Recommendation with Personalized User Interest and Time-aware News Popularity

Tao Qi, Fangzhao Wu, Chuhan Wu, Yongfeng Huang

TL;DR

This paper introduces PP-Rec, a novel news recommendation system that tackles the cold-start and diversity challenges by jointly considering personalized user interests and the time-aware popularity of news articles. By integrating these two factors, PP-Rec aims to provide more accurate and engaging news recommendations, especially for new users.

What Problem Does This Paper Address?

Online news platforms face a constant deluge of information, making it difficult for users to discover relevant content. Traditional personalized news recommendation systems often struggle with two key issues:

  • Cold-start problem: For new users with limited interaction history, it's challenging to accurately infer their interests, leading to poor recommendations.
  • Diversity problem: These systems tend to recommend news similar to what users have already consumed, potentially creating filter bubbles and limiting exposure to new information.

The paper argues that popular news articles, often covering significant events, can attract a broad audience and are inherently diverse, making them a valuable signal to address these limitations.

Methods

PP-Rec proposes a framework that combines a personalized matching score with a news popularity score to rank candidate news.

  • Personalized Matching Score: This score captures a user's interest in a news article. It's derived from the relevance between a user's interest embedding and the news content embedding.
  • Knowledge-aware News Encoder: This module generates news content embeddings by leveraging both the text of the news title and any associated entities. It uses multi-head self-attention and cross-attention mechanisms to capture relationships between words and entities.
  • Popularity-aware User Encoder: This component addresses bias in user behavior data caused by news popularity. It learns user interest embeddings by considering both the content and popularity of clicked news, using a content-popularity joint attention network.
  • Time-aware News Popularity Predictor: This module estimates the popularity of a news article, considering its dynamic nature. It combines:
  • News Content: Information from the news content embedding.
  • Recency: How recently the news was published.
  • Near Real-time CTR (Click-Through Rate): This captures immediate user engagement, helping to overcome the impression bias present in raw view counts.
  • Personalized Aggregator: The final ranking score is a weighted combination of the personalized matching score and the news popularity score. The weights are dynamically adjusted based on the user's representation, giving more importance to popularity for cold-start users.

Datasets

The authors built two datasets for their experiments:

  • MSN: Collected from the Microsoft News website (October 19 - November 15, 2019).
  • Feeds: Collected from a commercial news feed in Microsoft (January 23 - April 23, 2020).

For both datasets, the last week/three weeks were used for evaluation, and the rest for training and validation.

Key Insights and Ideas

  • Popularity as a Signal for Cold-Start and Diversity: The core insight is that news popularity is not just about what's trending, but a rich signal that can help overcome the limitations of sparse user data (cold-start) and the tendency of recommender systems to produce homogenous outputs (diversity).
  • Time-Awareness is Crucial for Popularity: News popularity is not static. The paper emphasizes the need to model this dynamism by incorporating recency and near real-time CTR, rather than relying solely on historical view counts.
  • Mitigating Bias in User Behavior: User click data can be skewed by popularity. The "popularity-aware user encoder" is a clever way to disentangle genuine user interest from the influence of trending news.
  • Unified Framework for Popularity Prediction: Combining content, recency, and CTR in a single model for popularity prediction offers a more comprehensive and accurate estimation.

Key Results

  • Overall Performance: PP-Rec significantly outperforms both popularity-based and personalized news recommendation baselines across various metrics (AUC, MRR, nDCG@5, nDCG@10) on both MSN and Feeds datasets, with statistical significance (p < 0.001).
  • Cold-Start Users: PP-Rec shows substantial improvements for cold-start users, demonstrating its effectiveness when user history is limited.
  • Recommendation Diversity: PP-Rec consistently improves recommendation diversity, as measured by intra-list average distance and new topic ratio, compared to other personalized methods.
  • Ablation Studies:
    • Removing the news popularity score significantly degrades PP-Rec's performance, highlighting its importance.
  • Removing the personalized matching score also hurts accuracy, confirming the need for user interest modeling.
  • Ablation on the popularity predictor shows that news content, recency, and CTR all contribute significantly to accurate popularity prediction.

Limitations and Future Work

  • Dataset Construction: The paper notes that there isn't an off-the-shelf dataset with news popularity information, requiring them to build their own. This might limit direct comparison with future work on identical benchmarks.
  • Computational Cost: While not explicitly detailed as a limitation, the complexity of knowledge-aware encoders and attention mechanisms can imply higher computational requirements.
  • Future Work: The paper's success suggests that further exploration of integrating diverse signals (like user demographics, social context, or more sophisticated temporal dynamics) into news recommendation could be fruitful. The authors also implicitly suggest that their approach could be adapted to other domains where popularity and cold-start issues are prevalent.

Summary updated

BibTeX
@inproceedings{qi2021pp,
  author = {Tao Qi and Fangzhao Wu and Chuhan Wu and Yongfeng Huang},
  title = {{PP-Rec: News Recommendation with Personalized User Interest and Time-aware News Popularity}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.acl-long.424},
}
Figure