Predicting the Importance of Newsfeed Posts and Social Network Friends

Tim Paek, Michael Gamon, Scott Counts, David Maxwell Chickering, Aman DhesiAAAI 2010 · DOI

TL;DR

This paper tackles the "triage problem" of social media—the overwhelming flood of newsfeed updates—by using machine learning to predict which posts and friends a user actually finds important. By training SVM classifiers on social metadata and message text, the researchers achieved up to 85% accuracy in identifying important friends and 64% accuracy for newsfeed posts, significantly outperforming the standard reverse-chronological approach.

What Problem Does This Paper Address?

In 2010, Facebook was exploding in popularity, and users were beginning to experience "newsfeed fatigue." With the average user receiving over 1,000 items per week, the default reverse-chronological sorting meant that high-value content was often buried under trivial updates.

The authors identify a critical gap: while Facebook used basic heuristics (like comment counts) to highlight "Top News," it ignored the actual text of the messages and the deep historical relationship between users. This paper asks: Can we mathematically model "importance" to help users filter the noise?

Methods

The researchers developed a custom Facebook application called "Newsfeed Tagger" to collect ground-truth data from users. They then employed Linear Support Vector Machines (SVM) using the Sequential Minimal Optimization (SMO) algorithm to classify items as "Important" or "Not Important."

The Newsfeed Tagger application interface for rating newsfeed items (a) and friends (b).

The novelty lies in their extensive Feature Engineering, which categorized signals into three buckets:

  1. Social Media Properties: Metadata like post type (photo vs. link), comment/like counts, and interaction history.
  2. Message Text and Corpus: They analyzed a 2-3 month "corpus" of past interactions using n-grams and LIWC (Linguistic Inquiry and Word Count) to categorize emotional and thematic tone.
  3. Shared Background Information: They calculated the overlap in profile fields like hometown, religion, political views, and interests.

Datasets

The study involved 24 active Facebook users (each with at least 200 friends). Over one business week, they collected:

  • 4,989 newsfeed ratings: Users rated post importance on a 0-100 slider.
  • 4,238 friend ratings: Users rated how interested they were in a friend's daily life.
  • Metadata & Text: Programmatic data pulled via the Facebook Open Stream API.

Key Insights and Ideas

  • Chronology is a poor proxy for interest: The researchers found a Pearson correlation of only $r=0.01$ between the age of a post and its importance rating. As shown in the scatter plot below, importance ratings are distributed almost randomly across time, empirically proving that "newest" does not mean "best."
  • Text is King: Contrary to some social networking theories that focus purely on "graph" connections, the message text and corpus features were the most predictive. Even for predicting friend importance, analyzing what people said to each other was more effective than just looking at how often they interacted.
Distribution of newsfeed (a) and friend (b) ratings, and the correlation between post age and importance (c).
  • Importance is Subjective: In a small subset of 44 cases where multiple participants rated the same post, the ratings differed by an average of 41 points. This suggests that a "one-size-fits-all" algorithm will always struggle compared to personalized models.
  • The "Friend Inference" Boost: Since a system might not always know how much you like a friend, the authors built a "combined model" that first predicts friend importance and then uses that prediction as a feature to judge the importance of their posts.

Key Results

  • Friend Importance: The best model achieved 85% accuracy (a 25% error reduction over the majority-class baseline).
  • Newsfeed Post Importance:
    • Using all features (including the user's explicit friend rating): 69.7% accuracy.
  • Using the "Combined Model" (inferring friend importance automatically): 64.4% accuracy.
  • Personalization: For a single user with a high volume of ratings (400+), a personalized model reached 69.6% accuracy, a massive 34.6% error reduction over their specific baseline.

Limitations and Future Work

The primary limitation is the sample size; 24 participants is small for a machine learning study, though the number of individual ratings was substantial. The authors also noted that privacy restrictions limited the amount of historical data they could scrape.

For future work, the authors suggest moving toward fully personalized models that learn a user's preferences over time. They also aim to study "ranking" specifically—not just binary classification—to see if they can perfectly re-order a feed to match a user's ideal priority list.

Summary updated

BibTeX
@inproceedings{paek2010predicting,
  author = {Tim Paek and Michael Gamon and Scott Counts and David Maxwell Chickering and Aman Dhesi},
  title = {{Predicting the Importance of Newsfeed Posts and Social Network Friends}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7518},
}
Figure