Walking Through Twitter: Sampling a Language-Based Follow Network of Influential Twitter Accounts

Axel Bruns; Ben Thies; Cornelius Puschmann; Felix Victor M\"unch

REVIEW 4 major objections 6 minor 2 references

Walking Through Twitter: Sampling a Language-Based Follow Network of Influential Twitter Accounts

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A modified rank degree method, run with the free Twitter standard API, can sample the German Twittersphere's follow network so that it approximates the top 1 to 10 percent of influential accounts.

desk verdict A pragmatic sampling hack with real code and data, but the validation doesn't yet separate the walk from simply grabbing high-follower accounts. read the letter →

arxiv 1908.07788 v3 pith:WXVQCYMB submitted 2019-08-21 cs.SI

Felix Victor Münch , Ben Thies , Cornelius Puschmann , Axel Bruns This is my paper · ORCID

classification cs.SI

keywords subgraphsamplingrankdegreemethodTwitterfollownetworkinfluentialaccountsGermanTwitterspheresocialmediadataminingcommunitydetection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a modified walk-based sampling method, run through the free Twitter standard API, can retrieve a follow-network subset that approximates the top 1 to 10 percent of influential German-speaking Twitter accounts. The authors adapted the rank degree algorithm—a deterministic graph sampler that moves from an account to the friend with the highest follower count—and ran 200 parallel walkers for six months, collecting 937,809 accounts. They show that the sample is orders of magnitude more active and more followed than a random baseline, that it contains on average 40 percent of a random German account's friends, and that it reaches 85 percent of such accounts. If correct, this gives small research teams a low-cost way to map the influential backbone of a language-based Twittersphere and to study opinion leadership, agenda setting, and public sphere structure without full network access.

What carries the argument

The central object is the rank degree method, a deterministic walk-based graph sampling algorithm. In its original form, a walker repeatedly moves from the current node to the connected node with the highest degree, removes the traversed edge from the graph, and keeps the new node; multiple walkers collapse when they meet. The paper's adaptation makes this scheme work on directed, API-restricted Twitter networks: friends are ranked by follower count instead of total degree, only the last 5,000 friends of an account are fetched in a single API call, edges are burned per directed direction, and walker collapse is dropped in favor of parallel walkers that jump to fresh random seeds when no unburned edge remains. This mechanism is what converts a few local API lookups per account into a sample that, the paper argues, preserves the influential core of the network.

What would settle it

Run the adapted sampler on a network whose true influential users are known, such as a complete national follow network collected before the API closed, and check whether the sampled top accounts have higher overlap with the actual top 10 percent by in-degree, PageRank, or reach than a random sample of the same size. If overlap is no better than random, the method's central claim collapses. A partial and cheaper check is to redraw the coverage test against a fresh random sample of current German-using accounts rather than the 2016 benchmark.

Watch

Extended reading notes

Core claim

The central claim is that a practical adaptation of the rank degree method can approximate the most central and influential accounts of a language-based follow network using only cost-free Twitter standard APIs. In the adapted procedure, each walker starts from a random German-interface account, looks up its most recent 5,000 friends, chooses the friend with the highest follower count whose interface language is German and whose edge is not yet burned, and repeats from that friend; when stuck, the walker jumps to a new random seed. Because the walk always moves to a high-follower account and burns traversed edges, the sample accumulates a set of accounts that are followed by many others, i.e., the influential core. The paper reports that the resulting influencer sample of about 199,000 accounts (those with at least one incoming sampled edge) has tweet activity and follower counts orders of magnitude higher than a random baseline, covers about 40 percent of the friends of an average German-using account, and reaches 85 percent of a random test sample with more than one friend. From this, the authors conclude that the sample approximates the proverbial top 1 to 10 percent of influential accounts in the German Twittersphere, and they demonstrate that community detection on the sample's 3-core reveals distinct topical communities such as German politics, football, Austrian politics, and YouTubers.

Load-bearing premise

The load-bearing premise is that the adapted walk—using directed edges, follower-count ranking, 5,000-friend truncation, 200 non-collapsing walkers, and seed-jumping—still identifies the influential core as well as the original rank degree method did on undirected networks; the authors explicitly state this transfer still has to be tested on known networks.

Editorial extensions

If this is right

Researchers can map the influential core of any language-based Twittersphere with only free API access, making cross-national comparison of public spheres feasible for small teams.
The sampled account set can serve as a monitoring frame for studying opinion leadership, agenda setting, automated accounts, and the behaviour of the most visible Twitter users over time.
The same walk can be redirected by changing the seed pool, for example to topic-based or event-based follow networks rather than language-based ones.
Because the sample covers 40 percent of an average account's friends and reaches 85 percent of a random test sample, it may serve as a proxy for the content-exposure backbone that mention or retweet networks miss, including silent listening.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The evidence compares the sample against a 2016 benchmark, so the top-10-percent claim is anchored to a network that has aged; a replication against a freshly drawn random sample of current German-using accounts would test the claim under present conditions.
Because the API returns the 5,000 most recent friends, the walk may favour recently active accounts; weighting coverage by friend-list position or timeline exposure could reveal whether the sample truly captures the most-seen accounts or merely the most recently active ones.
If the adaptations preserve centrality, the algorithm becomes a generic API-budget-aware network sampler whose parameters (walker count, friend-list truncation, ranking heuristic) could be tuned against known graphs, a direction the paper itself flags for future work.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Felix, quick take on arXiv:1908.07788. This is a useful paper for anyone who needs to collect a sample of influential Twitter accounts with only the free API. The authors adapt rank-degree sampling to a directed network and a rate-limited environment: they truncate friend lists at 5,000, rank friends by follower count, run 200 walkers, and use a TrISMA-derived seed pool. They release the code and present the first map of the German Twittersphere's influential core. The community structure (politics, sports, YouTubers, Swiss/Austrian clusters) looks plausible and matches what you'd expect.

What they do well: they are honest. The Outlook explicitly says the method 'still has to be tested with known networks' and that the directed application and non-dynamic ranking 'might lead to significant differences in the sample quality.' That is the key sentence in the paper.

The soft spot is central: the evaluation cannot distinguish the rank-degree walk from a simple rule that always moves to the highest-follower account. The sample's accounts have high follower numbers, high activity, 40% coverage of a random test sample's friends, and 85% reach. But any set of most-followed accounts would look similar, because preferential attachment makes those accounts widely followed. The baseline is a random sample of the same size; it isn't the right control. To support the claim that the walk preserves topology or adds value over a follower-count ranking, they'd need to compare against, say, the top 199k accounts by follower count, or run the method on a network they can fully observe. The current evidence is consistent with the weaker claim: this is a cheap way to collect a high-follower subset.

There are also smaller issues. The full sample is 938k accounts, but the quality statistics are computed on the ~199k accounts with at least one incoming edge. That's defensible for defining an 'influencer sample,' but it should be flagged more prominently. The community detection is a nice demo, not a validation.

None of this is fatal for the paper's practical purpose. If you want a low-cost, API-compliant way to get a set of influential German Twitter accounts and a bird's-eye view of their communities, this is a working tool. But the title and abstract claim more than the evidence supports. I'd want a revision that either provides the high-follower baseline or explicitly narrows the claims to what is shown.

Verdict: worth a serious referee, with a request for that comparison. The authors are clear thinkers and the code/data are real. I'd bring it to a reading group if anyone is thinking about Twitter sampling.

Referee Report

4 major / 6 minor

Summary. The paper adapts the rank degree graph-sampling method to the Twitter standard API in order to collect a sample of the German-speaking Twittersphere's follow network. The collection ran from December 2018 to May 2019 and gathered 937,809 accounts by running 200 parallel walkers that, at each step, select the friend with the highest follower count among the last 5,000 friends of the current account, subject to a German interface-language filter. The authors then define an 'influencer sample' of about 199,000 accounts with at least one incoming edge in the collected network and evaluate this subsample against a random baseline from the 2016 TrISMA collection. They report that accounts in the influencer sample have much higher activity and follower counts, that on average 40 percent of a random German Twitter account's friends are in the sample, and that the sample reaches 85 percent of accounts in a random test sample. A further test case applies infomap community detection and keyword extraction to the 3-core of the sampled network and identifies interpretable topical communities (e.g., German politics, Swiss politics, gaming, football, hard-right discourse). The authors conclude that the adapted method approximates the top 1 to 10 percent of influential accounts in the German Twittersphere and opens avenues for comparative research on other language-based Twitterspheres.

Significance. If the central claim holds, the paper offers a practical, low-cost method for mapping the influential backbone of a language-based Twittersphere under increasingly restrictive API conditions, which would be of clear value to computational social science and media studies. The manuscript has notable strengths: the method is described in enough detail to be replicated; the prototype code is openly available; the authors are transparent about the adaptations made to the original rank degree method; and the community-detection test is a concrete demonstration of how the data can be used. The observed differences between the influencer sample and the random baseline are large and visually striking. However, the current evaluation does not yet establish that the adapted walk preserves the topology of the influential core in a way that a simpler high-follower selection procedure would not, and the paper's own Outlook concedes that the centrality-preserving properties of the original method still need to be tested on known networks.

major comments (4)

[Evaluating the Sample Quality (Figure 5)] The comparison of follower counts between the influencer sample and the TrISMA benchmark is partly circular: the samplers explicitly choose the friend with the highest follower count at each step (Section 'Our adaptation and implementation of the Rank Degree method', step 3), so the follower-count distribution of the resulting sample is a direct consequence of the selection rule rather than an independent validation of the method's ability to find influential accounts. An additional baseline that selects the 199,000 accounts with the highest follower counts directly, or a random-walk variant that does not rank by followers, would be needed to isolate what the rank-degree walk contributes beyond simple follower-count maximization.
[Coverage and Reach (Table 1, Figures 6-11)] The coverage and reach metrics are compared only against a random baseline, not against a baseline consisting of the most-followed accounts. Because random Twitter users preferentially follow high-follower accounts, a sample constructed purely by picking high-follower accounts would also be expected to show elevated coverage and reach relative to a random sample. The reported 40 percent coverage and 85 percent reach are therefore consistent with the much weaker claim that the most-followed accounts in a language community are widely followed; they do not by themselves demonstrate that the walk-based sampling preserves influential-spreader topology or that the method outperforms a simple top-follower list.
[Outlook] The paper explicitly states that the adapted method 'still has to be tested with known networks' and that the application to directed networks and the non-dynamic ranking 'might lead to significant differences in the sample quality.' This is an admission that the load-bearing transfer from the original rank degree method, whose centrality-preserving properties were established in the lab, has not been validated for the Twitter adaptation. None of the current evaluation substitutes for that ground-truth test, since all metrics are computed either on the sampled network itself or against random baselines from a historical dataset.
[Evaluating the Sample Quality (Figures 4 and 5)] The benchmark TrISMA data were collected in 2016, while the sample's activity and follower counts were measured in 2018-2019. Follower counts generally accrue over time for existing accounts, and the activity measure is computed over different observation windows, so part of the reported 'orders of magnitude' difference may be a temporal artifact. The coverage and reach analysis correctly retrieves contemporaneous friend lists for the test sample, and a similar approach should be used for the activity and follower-count comparisons, or the temporal mismatch should be explicitly quantified and controlled.

minor comments (6)

[Abstract and Conclusion] The abstract and conclusion refer to 'the sample' when the quality metrics are computed on the filtered influencer sample of about 199,000 accounts, while the raw collection contains 937,809 accounts; the scope of the claim should be stated more precisely in both places.
[Figures 4 and 5] The phrase 'orders of magnitude' is used without a precise quantitative statement; reporting median or mean fold-differences and their confidence intervals would make the claim more informative.
[Coverage and Reach (Table 1)] The differences between the influencer and baseline samples are reported without confidence intervals or significance tests; given the large observed gap, adding a simple two-sample test or bootstrap intervals would strengthen the presentation.
[Sampling (Adaptation 1)] The decision to look up only the last 5,000 friends means that high-follower accounts followed earlier are systematically excluded; this is a further deviation from the original rank degree method and should be discussed explicitly as a potential source of bias.
[Test Case: Topical Communities in the German Twittersphere] The thresholds used for the 3-core filter, the top-50 keyword cutoff, and the 5-percent keyword-usage cutoff are stated to be arbitrary; the robustness checks mentioned in the text are not shown, so the relevant sensitivity analysis should be reported in the main text or the supplement.
[Figures 10 and 11] The term 'rank-reach' is used in the figure captions but is not defined in the main text; a one-sentence definition in the caption or the methods section would improve clarity.

Circularity Check

2 steps flagged · score 5.0 of 10

The 'influencer sample' validation is partly self-referential: the sampler greedily maximizes follower count, and the paper then cites follower count (plus popularity-driven coverage/reach) as evidence that the sample is influential.

self definitional [Methods and Analysis, 'Our adaptation and implementation of the Rank Degree method', steps 2-3; and 'Activity and Centrality' evaluation.]
"2. Look up the last 5,000 friends of w and rank them by their follower count. 3. Choose the friend v with the highest follower count ... the usage of the follower number is only a heuristic for assessing influence, an approximation of the in-degree in our network of interest ... this sample exhibits indeed a high-influence profile in terms of activity and in-degree centrality (as measured by the follower numbers reported by the Twitter API)."

The sample is constructed by repeatedly following the highest-follower-count friend available (subject to language and unburned-edge constraints). The same follower count is then used as evidence that the sampled accounts are influential, so the high follower-number distribution is largely a property of the selection rule rather than an independent empirical finding. Coverage and reach are also popularity-driven: a random account is much more likely to follow an account with many followers, so the reported 40 percent coverage and 85 percent reach are expected consequences of selecting high-follower nodes, not independent confirmation that the walk-based method preserves the influential-spreader topology of the original rank-degree algorithm.
fitted input called prediction [Methods and Analysis, 'Coverage and Reach' (final paragraph before 'Test Case: Topical Communities').]
"our sample still represents less than 10 percent of this population. Taking this and everything above into account, we conclude that the influencer sample is a good approximation of the most influential core of the German-using Twittersphere."

The 'top 1 to 10 percent' claim is operationalized by comparing the sample's size to a population estimate, while 'everything above' includes the follower-count metric that the sampler was explicitly designed to maximize. The paper never validates against an external ranking of influential accounts or against the full follow network; the conclusion that the sample is the influential top decile is therefore a restatement of the sampling criterion combined with a size calculation, rather than an independent test of the sample's influence.

full rationale

The central sampling claim is partially circular. The algorithm's core step ranks friends by follower count and walks to the highest-follower account; therefore, showing that the sample has high follower numbers is showing the construction's own input. Coverage and reach, while not directly optimized, are strongly coupled to follower counts through Twitter's preferential-attachment dynamics, so they provide only weak independent support. The paper does contain genuinely independent evidence: the high activity distribution is not directly selected for, and the community-detection/keyword test demonstrates that the sampled network is interpretable and resembles the Australian analysis. The Outlook explicitly concedes that the adapted method 'still has to be tested with known networks' and that the directed application and non-dynamic ranking 'might lead to significant differences in the sample quality'; this is an important validation gap and a correctness risk, but it is not itself circularity. Self-citations to TrISMA (Bruns et al., 2016) and Münch (2019) provide the seed pool and a qualitative comparison, but they are not the load-bearing argument that the adapted sampler preserves centrality; the rank-degree provenance is from external authors (Voudigari, Salamanos et al.). Because one of the main validation metrics is built into the selection rule while other metrics and the practical community-mapping application remain independent, the appropriate circularity score is 5: partial circularity, not a fully forced result.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No fitted model parameters or invented entities. The method uses hand-chosen operational thresholds and several domain assumptions. The most load-bearing is that rank-degree's centrality preservation survives the directed, follower-ranked, API-limited adaptation; the authors flag this as untested.

free parameters (4)

5,000 friend list truncation
The algorithm only inspects the most recent 5,000 friends of each account; this hand-chosen limit affects which high-follower friends can be selected and is justified by API rate limits.
200 parallel walkers = 200
Chosen to maximize use of 12 API keys; far fewer than the thousands of walkers used in original rank degree tests, which may affect sample quality.
3-core filter for community detection
Communities are detected on the 3-core of the sample (66k nodes), chosen to remove star artifacts; this changes the community structure reported.
Keyword thresholds = top 50 keywords and 5% usage cutoff
Arbitrary thresholds for community labeling; the authors report the results were robust to changes.

assumptions (5)

domain assumption Rank degree sampling preserves influential spreaders when about 20% of the network is explored (Salamanos et al. 2017).
This motivates the whole approach; the paper assumes the property transfers to the modified directed version.
domain assumption Follower count is a valid approximation of in-degree and of influence in the German Twittersphere.
Used in the sampling step and in evaluation; based on Cha et al. 2010 and the authors' reasoning.
domain assumption Twitter interface language setting identifies the German-speaking Twittersphere.
Used to filter seeds and candidate friends; the paper acknowledges this became a private property and the method must switch to tweet language.
domain assumption The 2016 TrISMA collection is a near-complete benchmark of German-using Twitter accounts.
Used as seed pool and as baseline for activity and follower comparisons; collected in 2016, not contemporaneous with the 2018-19 sample.
domain assumption The random test sample of 597 accounts with retrievable friends represents German-using accounts in 2019.
Coverage and reach estimates rely on this; accounts that were deleted, protected, or had fewer than 2 friends were excluded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Walking Through Twitter: Sampling a Language-Based Follow Network of Influential Twitter Accounts." pith.science (2026). https://pith.science/paper/WXVQCYMB

@misc{pith2026190807788,
  author       = {Pith},
  title        = {Pith review of: Walking Through Twitter: Sampling a Language-Based Follow Network of Influential Twitter Accounts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXVQCYMB}},
  note         = {Machine review of arXiv:1908.07788}
}

read the original abstract

Twitter continuously tightens the access to its data via the publicly accessible, cost-free standard APIs. This especially applies to the follow network. In light of this, we successfully modified a network sampling method to work efficiently with the Twitter standard API in order to retrieve the most central and influential accounts of a language-based Twitter follow network: the German Twittersphere. We provide evidence that the method is able to approximate a set of the top 1 to 10 percent of influential accounts in the German Twittersphere in terms of activity, follower numbers, coverage and reach. Furthermore, we demonstrate the usefulness of these data by presenting the first overview of topical communities within the German Twittersphere and their network structure. The presented data mining method opens up further avenues of enquiry, such as the collection and comparison of language-based Twitterspheres other than the German one, its further development for the collection of follow networks around certain topics or accounts of interest, and its application to other online social networks and platforms in conjunction with concepts such as agenda setting and opinion leadership.

Figures

Figures reproduced from arXiv: 1908.07788 by the authors.

**Figure 1.** Sample size as measured by the total number of edges over time (18-12-17 – 19-05-28) Sampling Our sample was collected from the mid of December 2018 until the end of May 2019 ( [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

**Figure 2.** Our adaptation of the rank degree algorithm. The top panel represents the sample after every iteration, the bottom panel represents the underlying network without the removed edges. The example network is based on a student interaction network (Heidler et al., 2014), filtered for indegree > 3, as available from https://github.com/gephi/gephi/wiki/Datasets. Evaluating the Sample Quality As we do not possess knowledge… view at source ↗

**Figure 3.** Distribution of the date of the last status by accounts in our sample at the end of the network collection timeframe (May 2019) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Comparison of our sample (‘Sample’) with all accounts collected in 2016 by TrISMA (‘Benchmark 2016’) regarding the distribution of the statuses per day since account creation. Nevertheless, as depicted in [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Comparison of our sample with all accounts collected in 2016 by TrISMA regarding the distribution of the follower count at the time of the sample collection. The spike between 100 and 1000 accounts is caused by a fully connected bot-net. A similar picture is drawn if w…

**Figure 6.** Figure 6: Distribution of accounts in the test sample over the percentage of their friends that can be found in the influencer sample (filtered for in-degree >= 1, leaving 199,180 accounts) [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

**Figure 7.** Figure 7: Distribution of accounts in the test sample over the percentage of their friends that can be found in the baseline sample (199,180 accounts drawn randomly from German-using accounts in TrISMA collection) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: Rank-coverage distribution of accounts in the test sample with at least 2 friends for the influencer sample (filtered for in-degree >= 1, leaving 199,180 accounts) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: Rank-coverage distribution of accounts in the test sample with at least 2 friends for the baseline sample (199,180 accounts drawn randomly from German-using accounts in TrISMA collection) As a baseline we drew a random sample from the German-using accounts in the TrISM…

**Figure 10.** Figure 10: Rank-reach distribution of accounts in the influencer sample (filtered for in-degree >= 1, leaving 199,180 accounts) [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Rank-reach distribution of accounts in the baseline sample (199,180 accounts drawn randomly from German-using accounts in TrISMA collection) The categorical difference between our influencer sample and the baseline becomes even more clear when examining the rank-distr…

**Figure 12.** Figure 12: central communities in the 3-core of our sample network; coloured by largest communities detected with the infomap community detection algorithm; node size represents Page Rank (Brin & Page, 1998); layout done with Force Atlas 2 in Gephi (Bastian et al., 2009) (colour…

**Figure 13.** Figure 13: Community graph of communities in the 3-core of our sample with over 300 accounts, at least 80 active accounts during the examined time frame, and edges with a weight of at least 150; edge width represents weight; edge direction follows clockwise curvature; edges colo…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[5]

If all vs in the seeds are leaves (degree = 1), select new initial seeds. 6. Repeat 1. – 7. until sample is of size x. From the steps outlined above, it becomes evident that the process is deterministic in that it completely depends on the initial seeds chosen (and, in case 5. is executed, on the new initial seeds). Moreover, any selected node can be visi...

work page 2016
[7]

Feuilleton

Repeat steps 2. - 7. with v as the new starting point. To use the full capacity of the API calls available to us, we used 200 parallel walkers. Contrasting to the original rank degree method, we do not let the walkers collapse when they land on the same node for time efficiency reasons but let them execute consecutively. While close to the feasible maximu...

work page arXiv 2016

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [5]

If all vs in the seeds are leaves (degree = 1), select new initial seeds. 6. Repeat 1. – 7. until sample is of size x. From the steps outlined above, it becomes evident that the process is deterministic in that it completely depends on the initial seeds chosen (and, in case 5. is executed, on the new initial seeds). Moreover, any selected node can be visi...

work page 2016

[2] [7]

Feuilleton

Repeat steps 2. - 7. with v as the new starting point. To use the full capacity of the API calls available to us, we used 200 parallel walkers. Contrasting to the original rank degree method, we do not let the walkers collapse when they land on the same node for time efficiency reasons but let them execute consecutively. While close to the feasible maximu...

work page arXiv 2016