[Paper Review] Modeling news spread as an SIR process over temporal networks

Elisa Mussumeci, Flávio Codeço Coelho|arXiv (Cornell University)|Dec 14, 2016

Opinion Dynamics and Social Influence6 references3 citations

TL;DR

This paper proposes a method to reconstruct news spread networks using semantic similarity and temporal proximity, modeling the dynamics as an SIR epidemic process. It demonstrates that real-world news dissemination closely follows SIR model predictions, enabling accurate prediction of media reach and detection of viral content.

ABSTRACT

News spread in internet media outlets can be seen as a contagious process generating temporal networks representing the influence between published articles. In this article we propose a methodology based on the application of natural language analysis of the articles to reconstruct the spread network. From the reconstructed network, we show that the dynamics of the news spread can be approximated by a classical SIR epidemiological dynamics upon the network. From the results obtained we argue that the methodology proposed can be used to make predictions about media repercussion, and also to detect viral memes in news streams.

Motivation & Objective

• To model the spread of news in digital media as a temporal network process driven by semantic similarity and publication timing.
• To test whether the spread of news articles follows the dynamics of the classical SIR epidemiological model on reconstructed networks.
• To develop a methodology for predicting media reach and identifying viral news based on early spread patterns.
• To explore the potential of network-based dynamics as a signal for detecting media manipulation or fake news.

Proposed method

• Used a skip-gram word embedding model trained on 2.6 million news articles to generate dense vector representations capturing semantic similarity.
• Computed document vectors as TF-IDF-weighted sums of word vectors to represent each article in a 300-dimensional semantic space.
• Defined article similarity using cosine distance between document vectors, with a threshold of ρ = 0.8 to establish influence links.
• Established temporal constraints by requiring influencer articles to precede target articles by less than 168 hours (7 days), based on lag distribution analysis.
• Reconstructed a directed temporal network of influence where edges represent likely information propagation from earlier to later articles.
• Simulated SIR dynamics on the reconstructed network using a transmission rate λ ∈ [0.00002, 0.00003] calibrated to match the empirical peak of article publication.

Experimental results

Research questions

• RQ1 Can news spread in digital media be accurately modeled as an SIR process on a dynamically reconstructed temporal network?
• RQ2 What combination of semantic similarity and temporal proximity best identifies causal influence between news articles?
• RQ3 To what extent does the empirical spread of news mirror the theoretical predictions of the SIR model in terms of peak timing and cumulative reach?
• RQ4 Can the reconstructed network and SIR dynamics be used to predict the final reach and duration of media coverage for a news event?

Key findings

• The reconstructed influence network achieved a giant component containing at least 80% of articles when similarity threshold ρ was set to 0.8.
• 95% of influencer articles were published within 7 days prior to the influenced article, validating a 168-hour (7-day) temporal window for influence.
• The simulated SIR model with λ = 0.0000215 closely matched the empirical distribution of article infections over time, as confirmed by 10,000 simulation runs.
• The peak number of infected articles in the simulation fell within the empirical range when λ was set between 0.00002 and 0.00003, indicating strong model fit.
• The study demonstrates that real-world news spread dynamics are well-approximated by the classical SIR model on a semantic-temporal network.
• The methodology enables prediction of media impact and detection of viral content based on early spread patterns, with potential for identifying manipulated or fake news.