Source-linked AI summary
The Simple Rules of Social Contagion
Nathan O. Hodas, Kristina Lerman
TL;DR
The paper asks why repeated information exposures on social media do not follow the independent cascade model’s pathogen-like prediction. It compares Twitter and Digg with a model that separates interface visibility from social influence, finding that accounting for visibility yields monotonically increasing response with exposure while explicit friend-vote feedback amplifies responses.
Problem
The paper addresses the gap between pathogen-like contagion models and empirical evidence that repeated social-media exposures can produce more complex responses.
Method
The authors compare Twitter and Digg data, model exposure visibility and social feedback, and use the model to forecast user activity.
Results
After visibility is accounted for, social contagion operates as simple contagion, with infection probability increasing monotonically with exposures; Digg’s explicit social feedback amplifies response.
Takeaways & Limitations
Cognitive limits and interface design shape information spread, so highly connected users can inhibit rather than amplify social contagion.
Abstract
from arXiv · showhide
It is commonly believed that information spreads between individuals like a pathogen, with each exposure by an informed friend potentially resulting in a naive individual becoming infected. However, empirical studies of social media suggest that individual response to repeated exposure to information is significantly more complex than the prediction of the pathogen model. As a proxy for intervention experiments, we compare user responses to multiple exposures on two different social media sites, Twitter and Digg. We show that the position of the exposing messages on the user-interface strongly affects social contagion. Accounting for this visibility significantly simplifies the dynamics of social contagion. The likelihood an individual will spread information increases monotonically with exposure, while explicit feedback about how many friends have previously spread it increases the likelihood of a response. We apply our model to real-time forecasting of user behavior.
I. DATA SOURCES
The study compares information spread on Twitter and Digg, using each platform’s user interface and social graph to track exposures and responses.
- Twitter presents friends’ tweets in a stream, with the most recent tweet or retweet at the top.
- Digg presents stories recently submitted or voted for by friends in chronological order, with the newest recommendation at the top.
- The Twitter dataset contains 2K URLs retweeted 213K times across a graph with almost 700K nodes and over 36M edges.
- The Digg dataset contains 3.5K promoted stories and voting activity from 140K users in a graph with 280K users and 1.7M links.
II. RESULTS
The analysis models repeated exposures together with message visibility and social feedback. After accounting for interface-driven visibility, contagion becomes simpler, while Digg’s explicit friend-vote signal amplifies responses.
- Repeated exposures initially raise infection probability but later saturate on Twitter or suppress further infection on Digg when aggregated across users.These aggregate patterns conflate users with different cognitive loads and information volumes.
- Message visibility is the probability that a user discovers a specific exposure, and it decays rapidly as newer messages push that exposure down the stream.The model measures visibility with the time response function T(∆t, nf).
- The multiple-exposure model combines the probability of finding exposures at their arrival times with a social enhancement factor for observed friend recommendations.Vn captures exposure visibility, while F(ne) represents social feedback associated with the number of recommending friends.
- On Twitter, exposures can be discovered independently because each message starts at the top of the stream, so any exposure may cause infection.
- Digg’s fixed first-recommendation position makes additional votes visible as a friend-count badge, and this social feedback produces large amplification of infection probability.
- The model forecasts whether users retweet or vote within the subsequent 30 seconds by comparing predicted and observed probabilities.
- Rapid visibility decay and lower susceptibility among highly connected users help explain why information spreads less widely than predicted by the independent cascade model.The analysis is content agnostic and removes most spam; novelty decay cannot be evaluated for Digg because observation ends before front-page promotion.
III. CONCLUSIONS
The paper shows that cognitive limits make information spread differ fundamentally from disease-like contagion. Comparing sites and accounting for visibility isolates a simpler exposure-response pattern.
- Cognitive constraints in processing information produce important differences between social and pathogenic contagion.
- Highly connected users can inhibit information spread because information overload reduces visibility and increases the social signal needed to respond.Visibility decays faster as the number of followed friends increases, reducing susceptibility to any single exposure.
- Accounting for visibility makes social contagion behave as simple contagion, with response probability increasing monotonically across exposures.The model separates interface visibility from social influence using comparisons across websites.
A. Methods
The methods estimate message visibility and exposure-response probabilities from URL-sharing activity on social media. Likelihood fitting and empirical calibration then support response modeling and forecasting, subject to an event-count probability floor.
- The data are decomposed into user-specific time series for URL-containing tweets or votes, with visibility calculated at one-second intervals.Visibility measures the probability of finding a URL in the stream at a given time after exposure.
- Visibility is modeled through time response functions that depend on elapsed time and the user’s number of friends.Twitter uses single-exposure events for this calculation, while Digg relaxes that constraint because such events are scarce.
- Model calibration compares calculated and observed probabilities and minimizes weighted mean absolute percent error while fitting visibility and friend-count functions.The fitted forms and parameters differ between Digg and Twitter.
- The exposure model estimates baseline response probability and multi-exposure enhancement by maximizing likelihood across exposure counts.For one exposure, F(1) is fixed at 1 to estimate P(ν); higher-exposure values estimate F(ne) conditional on that baseline.
- Forecasted probabilities are bounded below by the number of observed events, producing apparent deviations at very high and low predicted probabilities.The probability floor rises above the predicted-equals-observed line because extreme visibility and influence conditions are uncommon.
B. Approximating Visibility Functions
The model separates the probability that users see URL-containing messages from the social enhancement produced by multiple exposures. For Twitter, an approximation replaces detailed exposure combinations with a visibility-based factor; for Digg, the interface makes only one term relevant.
- The infection probability depends on both the probability of seeing a URL n times and an enhancement factor f(n; nf) for multiple exposures.The enhancement factor captures the collective effect of multiple exposures, while visibility determines which messages are observed.
- For Digg before front-page promotion, the interface orders URLs by the first recommendation, so only one explicit-observation term contributes.Approximating f(ne; nf) as the social enhancement factor F(ne) yields the probability of Digging a received URL.
- For Twitter, each tweet’s discovery probability is P(nf)T(t−ti, nf), and multiple tweets containing the same URL can appear in chronological order.Each message’s visibility decays according to its arrival time in the user’s stream.
- Enumerating Twitter’s exact observation probabilities becomes computationally prohibitive because Vn contains ne!/n!(ne−n)! combinations.The paper therefore proposes an approximate form for Twitter’s contagion probability as ne grows.
- Generating functions represent the many Vn terms compactly and support an approximation that uses a single social enhancement factor Ftw(ne) for seeing any URL.The approximation is evaluated by comparing it with the exact expression through a time-dependent ratio F*(ne,t).
- When visibility is at either extreme, F* tends toward α+β and varies weakly with time, supporting the approximation of Pexact by P*.The two limiting cases use τ0=τmax and τ0=vmin.