Source-linked AI summary
Twitter reciprocal reply networks exhibit assortativity with respect to happiness
Catherine A. Bliss, Isabel M. Kloumann, Kameron Decker Harris, Christopher M. Danforth, Peter Sheridan Dodds
TL;DR
The paper asks whether happiness is assortative in Twitter social subnetworks, where disentangling contagion from homophily remains a challenge. It constructs time-bounded reciprocal-reply networks and finds that users’ average happiness correlates positively with that of users up to three links away.
Problem
Whether happiness assortativity appears in Twitter subnetworks is an empirical question, while distinguishing contagion from homophily remains difficult.
Source: §1 · p. 1Method
The authors construct time-bounded reciprocal-reply networks and compare linked users’ happiness, including correlations across one-, two-, and three-link neighborhoods.
Source: §1 · p. 2, §2.3 · p. 6Results
Users’ average happiness is assortative in reciprocal-reply networks, with correlations detectable up to three links away.
Source: §4 · p. 9Takeaways & Limitations
The findings provide evidence of a social subnetwork within Twitter and support treating it as a social service through time-bounded interactions.
Source: §4 · p. 9Because the study uses a tweet subsample, missing data may affect inferred network topology, and further work is needed to establish those effects.
Source: §4 · p. 9
Abstract
from arXiv · showhide
The advent of social media has provided an extraordinary, if imperfect, 'big data' window into the form and evolution of social networks. Based on nearly 40 million message pairs posted to Twitter between September 2008 and February 2009, we construct and examine the revealed social network structure and dynamics over the time scales of days, weeks, and months. At the level of user behavior, we employ our recently developed hedonometric analysis methods to investigate patterns of sentiment expression. We find users' average happiness scores to be positively and significantly correlated with those of users one, two, and three links away. We strengthen our analysis by proposing and using a null model to test the effect of network topology on the assortativity of happiness. We also find evidence that more well connected users write happier status updates, with a transition occurring around Dunbar's number. More generally, our work provides evidence of a social sub-network structure within Twitter and raises several methodological points of interest with regard to social network reconstructions.
1. Introduction
The paper argues that Twitter’s follower ties and one-way replies do not establish interaction, and proposes reciprocal replies as a basis for studying its social network. It examines happiness assortativity in time-bounded networks, including correlations with users up to three links away, and uses a topology-preserving null model to assess network structure’s role.
1. Introduction
- Follower links and one-way replies do not establish that users actually interact or that a reply’s recipient considered it.Broadcasting reaches followers without guaranteeing they read or respond; a one-way reply provides evidence of receipt, but not reciprocity.Source: §1 · pp. 1–2
- Two users are linked in the proposed network only when each has replied to the other at least once during the chosen period.The paper treats following and unreciprocated replies as insufficient evidence of interaction.Source: §1 · p. 2
- Stale ties and long-term averaging can obscure current interactions and changes in users’ happiness, motivating day-, week-, and month-scale networks.The authors note that accumulated links can inflate network statistics and blur the distinction between repeated interactions and a single recent interaction.Source: §1 · p. 2
- The study tests whether happiness is assortative in reciprocal-reply networks and whether correlations extend across multiple links, amid difficulty distinguishing contagion from homophily.Prior claims of contagion have been criticized because observed similarity may be difficult to separate from homophily.Source: §1 · pp. 1–2
- To assess network topology’s role in happiness correlations, the authors permute happiness scores across nodes while preserving network structure.They also examine users’ average happiness and that of their neighbors one, two, and three links away.Source: §1 · pp. 2–3
2. Methods
The methods build time-bounded reciprocal-reply graphs from sampled Twitter messages, characterize their topology, and assign users frequency-weighted happiness scores for neighborhood comparisons.
2.1. Data
- From September 2008 to February 2009, researchers retrieved over 100 million tweets, estimated 38% coverage, and analyzed nearly 40 million message-reply pairs using API reply metadata.The API provided message and user identifiers for tweets made with its built-in reply function.Source: §2.1 · p. 3
- Only built-in replies with API-linked message and user IDs entered the network; handle-only mentions were excluded, and missing replies could inflate inferred path lengths.Unobserved replies could also leave users unconnected, while interactions that did not use the built-in reply function were not captured as replies.Source: §2.1 · p. 3, Figure 3 · p. 3
2.2. Reciprocal-reply network
- Each time-bounded graph treats users as nodes and reciprocal replies as links, removes self-loops, and tracks node, degree, clustering, component, and giant-component statistics across daily, weekly, and monthly windows.Clustering is calculated using the number of triangles and connected triples.Source: §2.2 · pp. 3–4, Figure 6 · p. 6
- Degree assortativity uses Spearman correlation of endpoint degree ranks; the analysis also considers user pairs at shortest path lengths of two and three.The authors choose Spearman over Pearson because it assesses rank similarity and is less sensitive to outliers.Source: §2.2 · p. 4
2.3. Measuring happiness
- User happiness is a frequency-weighted average of labMT word scores across authored tweets, excluding near-neutral stop words.The word scores were based on human ratings; words within one point of 5 were excluded, leaving a 3,686-word list.Source: §2.3 · pp. 4, 6
- Spearman correlations compare connected users’ happiness and extend to users two and three links away.The score for each user is based on their authored tweets during the time period being analyzed.Source: §2.3 · pp. 4, 6
3. Results
Users’ average happiness is more similar to nearer network neighbors, and the association weakens with distance. Network-preserving randomization and a word-bag comparison provide further checks on the pattern, alongside results on network structure and user degree.
3.1. Reciprocal-reply network statistics
- kmin ≈ 34 and α = 3.5; for the sample week, the Kolmogorov-Smirnov test found no evidence against a power law (D = 2.28 × 10−2, p = 0.095, n = 203852).The sample month had the same exponent and statistically stronger evidence of a power law.Source: Figure 7 · p. 6, §3.1 · p. 7
- Users increased over the study while average degree, degree assortativity, and giant-component share stayed fairly constant; clustering declined, likely as more nodes reduced closed-triangle share.Maximum-degree fluctuations reflected bursts of high-volume reply exchanges by celebrities or companies.Source: §3.1 · pp. 6–7
3.2. Measuring happiness
- p < 10^-10 for all week correlations; happiness-happiness correlations weakened across one-, two-, and three-link neighborhoods.Users’ happiness was more similar to nearer neighbors than to users two or three links away.Source: §3.2 · p. 7, Figure 10 · p. 8
- Average happiness rises gradually with user degree; the authors say this may support an upper limit on active interactions near Dunbar’s number.Large-degree users were relatively few, and their word use differed from that of smaller-degree users.Source: §3.2 · p. 8, Figure 12 · p. 8
3.3. Testing assortativity against a null model
- 100 happiness-score permutations preserving network topology yielded no statistically significant relationship in a sample week.The null model retained the one-, two-, and three-link network structures while permuting users’ happiness scores.Source: §3.3 · p. 9, Figure 13 · p. 9
- Di,j = .267 observed versus Di,j = .167 randomized mean word-bag similarity provided no evidence that happiness correlations reflect similar word usage.Source: §3.3 · p. 9
4. Discussion
The discussion presents reciprocal-reply networks as a dynamic social subnetwork and reports happiness assortativity extending three links. It notes that missing data and the unresolved distinction between homophily and contagion constrain interpretation.
4. Discussion
- Reciprocal replies provide a dynamic Twitter subnetwork: week-scale links reflect current activity rather than ties that remain indefinitely.The authors argue that millions of weekly replies and observed happiness assortativity support Twitter’s use as a social service.Source: §4 · p. 9
- Across 25 weeks, node counts rose and clustering declined slightly, while giant-component share and Spearman degree assortativity remained fairly constant.The authors note that the clustering decrease may reflect subsampling or growth in network size; Spearman is less sensitive than Pearson to extreme degrees.Source: §4 · p. 9
- The analysis uses a subsample of tweets, so missing data may affect inferred network topology.The authors call for further work on the consequences of missing messages; subsampled scale-free networks are not necessarily scale-free.Source: §4 · p. 9
- Happiness assortativity is detected up to three links away, and the finding does not appear to reflect users tweeting similar words.The reported correlations are smaller than those in Bollen et al.’s six-month reciprocal-follower network, a difference the authors consider likely related to methodological choices.Source: §4 · p. 9
- The paper does not separate homophily from contagion, leaving their respective roles in the observed assortativity unresolved.The authors suggest reciprocal-reply networks as a framework for future study of these effects, while noting that missing data can introduce errors.Source: §4 · pp. 9–10
Appendix
The appendix documents network distributions, assortativity measures, visualizations, word comparisons, and data coverage. Together, these materials describe the checks and displays accompanying the study’s network analyses.
Appendix
- D = 1.82×10−2 (p = 0.35, n = 495881) provides no evidence against a power-law fit to the January 2009 degree distribution (α = 3.50, kmin = 109).The fit uses the procedure of Clauset, Shalizi, and Newman [41].Source: Figure A1 · p. 12
- Pearson degree correlation is more sensitive to extreme values and obscures rank-degree assortativity, so the study uses Spearman correlation.For happiness assortativity across labMT thresholds, Spearman also exhibits less variability than Pearson; at ∆h = 0, variation is lower across thresholds.Source: Figure A2 · p. 12, Figure A3 · p. 13
- The Week 1 reciprocal-reply network visualization sizes nodes by degree and colors them by detected degree grouping; the daily visualization colors detected communities.The weekly figure uses Gephi’s implementation of the algorithm suggested by Blondel et al. [40], while the daily figure cites [49].Source: Figure A4 · p. 14, Figure A6 · p. 16
- In the Week 1 happiness visualization, node colors encode happiness for users with more than α = 50 labMT words, and similar colors visually illustrate assortativity.Those users comprise 57% of the week’s nodes; the visualization clusters nodes according to connectivity.Source: Figure A5 · p. 15
- Observed connected-user word-bag similarity averages 0.267, versus 0.167 after random permutation; most observed pair scores are below 0.4.The randomized value is nonzero because users share English, and the authors describe the observed increase as small.Source: Figure A8 · p. 18
- Tables A1 and A2 list the 150 most frequent labMT words, respectively excluding and including stop words.Table A3 summarizes weekly network statistics, while Table A4 reports observed-message coverage and reply-message counts; the observed share falls as total tweets grow.Source: Table A1 · p. 19, Table A2 · p. 20, Table A3 · p. 21