Source-linked AI summary
Prioritizing Original News on Facebook
Xiuyan Ni, Shujian Bu, Igor L. Markov
TL;DR
Original news is a valued dimension of news quality, yet highly original stories are rare and news posts gain their greatest exposure early. The paper estimates originality using citation-based PageRank normalized by news event, with clustering and hourly updates. After deployment in Facebook News Feed, exposure of original content grew and users received more content they liked.
Problem
Estimating originality is difficult because articles are hard to judge in isolation, while highly original news is rare in a large online inventory.
Source: §1 · p. 1, §3.1 · p. 3Method
The system estimates originality using article citation-graph PageRank normalized within news events, clustering related stories and refreshing the signal hourly.
Source: §1 · p. 1, §6.4 · p. 8, §7 · p. 9Results
After deployment in Facebook News Feed, exposure of original content grew, and users received more content they liked.
Source: §7 · p. 9Takeaways & Limitations
The authors report that the positive impact on the news ecosystem should, over a longer timeframe, encourage publishers to invest more in original content.
Source: §7 · p. 9The approach excludes scraped and spun content, which dedicated systems handle and which does not appear in the News Feed inventory considered.
Source: §3.1 · p. 3
Abstract
from arXiv · showhide
This work outlines how we prioritize original news, a critical indicator of news quality. By examining the landscape and life-cycle of news posts on our social media platform, we identify challenges of building and deploying an originality score. We pursue an approach based on normalized PageRank values and three-step clustering, and refresh the score on an hourly basis to capture the dynamics of online news. We describe a near real-time system architecture, evaluate our methodology, and deploy it to production. Our empirical results validate individual components and show that prioritizing original news increases user engagement with news and improves proprietary cumulative metrics.
1 INTRODUCTION
The paper treats originality as a consequential dimension of news quality, but original reporting is rare and difficult to identify amid online news. It examines news originality and its life cycle, then develops a scalable, near-real-time signal and evaluates it for News Feed ranking.
1 INTRODUCTION
- Original reporting initiates the news cycle and drives the news industry, while originality distinguishes reporting from duplicates, lightly edited text, and coverage referencing original news.The authors argue that prioritizing original news is in everyone’s long-term interest.Source: §1 · p. 1
- Highly original news is rare in a large inventory, and posts typically receive their greatest exposure in the first couple of hours before a long tail.These findings point to the need for prompt scoring to affect News Feed ranking.Source: §1 · p. 1
- The introduction illustrates how a news article can cite another publisher’s story, which in turn cites an article from its own publisher.This citation pattern motivates analyzing links among news articles.Source: Figure 1 · p. 1
- The authors find content alone insufficient to judge originality, so they also use behavioral signals such as citations of prior posts and filter articles linked to nefarious activity.They evaluate the signal offline against professional journalists’ ratings and online through an A/B test that monitors news ranking.Source: §1 · p. 1
- The system combines PageRank with three-step clustering to estimate originality at large scale, using hourly incremental updates to capture news dynamics.The authors identify PageRank normalization within clusters as particularly novel.Source: §1 · p. 1
2 BACKGROUND
The background introduces article-level citation graphs and PageRank, article representations for comparing text, and the stages of Facebook News Feed ranking. It situates the originality signal within that ranking system while noting that the approach is not specific to Facebook.
2.1 The News Citation Graph
- The paper models news articles as vertices and links between them as directed citation edges, then computes PageRank to estimate individual article credibility; earlier work applied PageRank to Web ranking [4] [5] [34] and news-source ranking [34].The paper treats disproportionate citation by peers as an indicator of higher journalistic credibility.Source: §2.1 · p. 2
2.2 The News Article Representation
- BERT provides article representations for comparing text, and Sentence-BERT uses a Siamese network architecture for semantic similarity estimation [28].BERT can be specialized for a use case by adding a dense layer and training on labeled data.Source: §2.2 · p. 2
2.3 News Feed Ranking
- Facebook’s News Feed ranks content through inventory, signals, prediction, and relevance scores, where predicted user actions contribute to each item’s final relevance score.The originality signal is deployed within this system, whose ranking changes are described as subtle.Source: §2.3 · p. 2, Figure 2 · p. 3
2. BACKGROUND
- News Feed ranking objectives include long-term user satisfaction, community considerations, and integrity, and the originality signal changes ranking within that broader system.Other social media platforms use conceptually similar ranking systems, so the paper’s contributions are not specific to Facebook.Source: §2.3 · p. 2, §2 · p. 2
3 PROBLEM ANALYSIS
The section organizes news by originality, from copied or lightly edited material to peer-recognized original reporting, and examines how stories gain and lose audience attention over time. It identifies constraints on judging originality and on when ranking interventions can affect engagement.
3.1 The Landscape of News Originality
- The originality spectrum ranges from scraped or spun content through low-effort and heavily derivative articles to peer-recognized original reporting.The categories also include potentially original work that lacks peer recognition.Source: §3.1 · p. 3
- Scraped and spun articles can eclipse original content and undermine its value; Table 1 illustrates spun content through isolated edits to another publisher’s articles.Scraped content is copied without editorial effort, while spun content makes minor modifications.Source: §3.1 · p. 3, Table 1 · p. 3
- Most news articles fall into the somewhat unoriginal bucket, whose articles may provide useful information while requiring little effort to produce.The figure estimates total views received by articles in each originality bucket.Source: §3.1 · p. 3, Figure 3 · p. 3
- Potentially original articles may lack peer recognition, so citation signals alone cannot distinguish them from unoriginal articles.Opinion pieces with little support often fall into this bucket.Source: §3.1 · p. 3
3.2 The Dynamics of Online News
- News engagement for the Spiderman story fell significantly on day two and vanished by September 29, three days after the story broke.Almost 800 websites covered the story on its first day.Source: §3.2 · p. 4
- Later information and increased article inventory did not produce another engagement peak for the Ebola vaccine story, and traffic nearly vanished after seven days.The World Health Organization’s approval two days after the initial report coincided with increased inventory, but traffic remained mostly flat.Source: §3.2 · p. 4
- Because ranking interventions can be effective only early in a story’s life-cycle, the study focuses on articles published within the previous seven days.The originality intervention does not change the existing post inventory.Source: §3.2 · p. 4
4 ESTIMATING NEWS ORIGINALITY
The method estimates news originality from contextual signals, especially peer recognition in a news citation graph, rather than relying on content alone. It normalizes global PageRank within news-event clusters and uses the resulting score in News Feed ranking, while peer citations create a delay for new articles.
4 ESTIMATING NEWS ORIGINALITY
- News originality is estimated from contextual signals because content analysis alone is neither sufficient nor necessary to assess it.The paper frames originality as both the process by which content is created and its quality.Source: §4 · p. 4
- Global PageRank on the news citation graph measures peer recognition, including articles cited across publishers or later news events.Global values are used because quality articles may cite out-of-topic background and receive citations under later events.Source: §4 · p. 4
- Within each news-event cluster, the method normalizes articles’ global PageRank scores, treating the highest values as most likely original.The parameter p defaults to 1.0; increasing it favors articles with higher global PageRank.Source: §4 · p. 5
- New articles cannot be evaluated for originality until peers cite them, creating a delay that motivates near real-time scoring.Source: §4 · p. 5
- The originality score is converted to P(original); only scores above the promotion threshold qualify, and P(original) enters relevance as a second-order term.Relevance also includes probabilities of comments, shares, and likes; originality is one ranking component among others.Source: §4 · p. 5
5 IMPLEMENTATION AND SCALING
The system computes PageRank on a recent news-citation graph, filters citations vulnerable to manipulation, and normalizes article scores within event clusters. It uses BERT-based title similarity and a three-step clustering approach designed to process large inventories efficiently.
5.1 PageRank with Integrity Considerations
- The citation graph links recent news articles through canonicalized URLs, and PageRank is computed after filtering citations vulnerable to link farming.Articles without incoming or outgoing citations are excluded from the computation.Source: §5.1 · p. 5
5.2 News Event Clustering
- PageRank is normalized within event clusters because articles in larger clusters can receive more citations and higher scores despite similar originality.The workflow first computes global PageRank values, clusters articles by event or story, and then normalizes scores within each cluster.Source: §5 · p. 5, §5.2 · p. 6
- A Siamese BERT network produces title embeddings, and three-step clustering combines text hashing, KNN, and greedy local search.The KNN graph uses 5,000 nearest neighbors per vector; local clustering addresses cases such as low-coverage local news that top-down clusters handle poorly.Source: §5.2 · p. 6, Figure 7 · p. 6
5.3 Scalability
- Distributed clustering supports scalability and surge robustness, but server capacity must balance enough work to justify distribution against the available RAM.The design aims to process snapshots of 2M–3M articles in minutes and can apply to other social-network platforms with larger inventories.Source: §5.2 · p. 6, §5.3 · p. 7
6 EVALUATION AND DEPLOYMENT
The evaluation tests embedding and clustering components, rates the originality signal with professional journalists, and examines its production use. A live A/B test found no statistically significant deterioration in proprietary metrics before the signal was enabled for most users.
6.1 Evaluation of Embeddings and Clustering
- Embedding training used 100K randomly sampled English news-title pairs, while a separate 10K article-pair sample evaluated clustering.The title-pair data were split 40% for finetuning, 10% for validation, and 50% for testing.Source: §6.1 · p. 7
- The BERTPairwise model consistently outperformed pre-trained state-of-the-art embeddings in ROC AUC evaluation.The comparisons used FastText and Pytorch-BigGraph embeddings and thresholds 0.5, 1.5, and 2.5; Table 3 presents the embedding comparison.Source: §6.1 · p. 8, Table 3 · p. 8
- The clustering evaluation compared predicted same-cluster labels against professional annotations on 10K English article pairs and assessed precision and recall against DBSCAN and Louvain.Table 4 presents the three-stage clustering comparison; the supplied text does not state its outcome.Source: §6.1 · p. 8, Table 4 · p. 8
6.2 Evaluation by Professional Raters
- 90% accuracy in identifying original news was reported for articles rated 3.0 by professional raters.The sample comprised the most-viewed articles identified as original and the most-viewed article not identified as original from the most-viewed news domains over seven days.Source: §6.2 · p. 8
6.3 An Illustrative Example
- In the Kobe Bryant crash example, more than 200 articles joined the cluster within 10 minutes, and the first report by TMZ ranked highest.The signal only prioritizes an original article when it is in the user's feed inventory; publisher preferences and friend affinity continue to affect what users see.Source: §6.3 · p. 8
6.4 Production Deployment and Evaluation
- The signal is recalculated hourly, but crawling and embedding-generation bottlenecks, along with the time needed for citations to appear, constrain early promotion.The authors say more frequent workflow runs could find and promote originals earlier, with further infrastructure optimization likely to help.Source: §6.4 · p. 8
- The live A/B test found no statistically significant deterioration in proprietary metrics; the signal was enabled for most users in English-language News Feed in June 2020.The test ran for several weeks, with prior ranking rules for the control group and revised rules for a small test group.Source: §6.4 · p. 8
- The authors identify possible publisher manipulation and note PageRank abuse protections and Facebook integrity monitoring focused on coordinated inauthentic behavior.Source: §6.4 · p. 8
7 CONCLUSIONS AND PERSPECTIVES
The paper proposes prioritizing original news in social networks by normalizing article PageRank within news events. It reports deployment in personalized Facebook News Feed and increased exposure of original content alongside more content users liked.
7 CONCLUSIONS AND PERSPECTIVES
- The strategy estimates originality by normalizing news-article PageRank scores within each news event.Source: §7 · p. 9
- After deployment in personalized Facebook News Feed, exposure of original content grew and users received more content they liked.The authors state that these developments should encourage publishers to invest more in original content over a longer timeframe.Source: §7 · p. 9