Pith. sign in

REVIEW 5 major objections 8 minor 45 references

Quantifying Polarization: A Comparative Study of Measures and Methods

T0 review · 5 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read No single scalar measure reliably captures political polarization from comment-leaning distributions, and an adaptation of the burst detection algorithm recovers the number of modes that scalar scores miss.

desk verdict A careful comparative review of polarization measures undercut by an unvalidated burst-detection method that is claimed as the headline contribution. read the letter →

arxiv 2501.07473 v1 pith:UZGK4AGG submitted 2025-01-13 cs.CY cs.SIphysics.soc-ph

Edoardo Di Martino , Matteo Cinelli , Roy Cerqueti , Walter Quattrociocchi This is my paper · ORCID
classification cs.CYcs.SIphysics.soc-ph
keywords politicalpolarizationburstdetectionmodeunimodalityYouTubecommentsmeasures2020U.S.presidentialelection
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

Polarization is often measured by how far an opinion distribution departs from unimodality, but the paper shows that five established measures — the Bimodality Coefficient, the dip test, Distance from Unimodality, Van der Eijk's A, and Balance — disagree substantially on the same data. On synthetic distributions and on 117,457 YouTube comment sections from the 2020 U.S. presidential election, the Bimodality Coefficient tends to over-flag skewed unimodal distributions as bimodal, while the other measures mostly classify comment sections as unpolarized. To provide shape information that scalar scores lack, the authors adapt a burst detection algorithm to count the modes of a leaning distribution, treating each user's leaning score as a timestamp and aggregating nearby bursts into single modes. A sympathetic reader would take the paper's central claim to be that this adaptation gives a scalable, automated way to recover the number of modes in large-scale political-discussion data, supplementing measures like the dip test and Distance from Unimodality.

What carries the argument

The central object is a repurposed burst detection algorithm for discrete timed events, supplemented by an aggregation rule. In the adaptation, each comment is an event whose 'timestamp' is the poster's inferred leaning; duplicate timestamps are resolved by rounding to three decimals and adding $\epsilon = 0.0001$. The algorithm returns an intensity level for each interval on the leaning axis, and the aggregation rule merges two same-level bursts when their gap is less than $\phi = k \cdot \mathrm{HDI}(\alpha)$, where HDI is the Highest Density Interval of the leaning distribution; a burst contained in a lower-level burst is removed, and any remaining burst with intensity $\ge 3$ is counted as a mode. This two-step pipeline is what converts a raw array of leaning scores into a mode count.

What would settle it

Take the 65,928 comment sections with at least 50 active users and run the burst pipeline with $\alpha = 0.05$, $k = 0.5$ and with $\alpha = 0.3$, $k = 0.25$; the adaptation's claim to be a reliable mode detector would be falsified if the two parameter settings produce materially different burst counts on a large share of sections that the dip test and DFU both score as clearly unimodal.

Watch

Extended reading notes

Core claim

The paper's central claim is that no single scalar measure reliably quantifies polarization as departure from unimodality on real social-media data, and that the missing piece — the shape of the leaning distribution, and in particular its number of modes — can be recovered by re-purposing a burst detection algorithm originally built for discrete event streams. On the YouTube data, the Bimodality Coefficient marks 78.42% of comment sections as bimodal, while the dip test, Distance from Unimodality (DFU), Van der Eijk's A, and Balance mostly place sections below polarization thresholds; the paper reads this disagreement as evidence that a polarization measure must be chosen with knowledge of the distribution's shape. The proposed method sorts comments by rounded user-leaning 'timestamps', runs the burst detector, merges bursts whose gap is below $\phi = k \cdot \mathrm{HDI}(\alpha)$, removes contained bursts, and counts bursts of intensity at least 3 as modes. The authors argue that this count aligns with the DFU and the dip test — many detected modes make a distribution look uniform and hence unpolarized — and that it adds interpretability to those scalar scores.

Load-bearing premise

The load-bearing premise is that the bursts the algorithm finds on a pseudo-time series of rounded leaning scores plus a tiny constant are the real modes of the underlying opinion distribution, and that the parameter set $s = 1.7$, $\gamma = 0.9$, $\alpha = 0.05$, $k = 0.5$ is valid without independent calibration.

Editorial extensions

If this is right

  • The adaptation gives a concrete way to scale mode counting: instead of inspecting individual distributions, a researcher can run the burst pipeline over tens of thousands of comment sections.
  • All five measures cannot agree on a single polarization verdict for the same collection; any study relying on one measure alone risks a conclusion that another measure would reverse.
  • Center-leaning videos tend to draw more polarized audiences under four of the five measures, suggesting that centrist comment sections act as a meeting ground for otherwise separated partisan groups.
  • Higher burst counts track lower DFU and dip-test scores, so the mode count and the scalar measures are complementary rather than redundant.
  • The near-universal single-burst finding (87% of analyzed sections) supports the paper's reading that most YouTube political discussions are unimodal and ideologically segregated, not two-sided.

Reading between the lines

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

  • A natural next test would be to validate the pipeline against hand-annotated mode counts or against mixture models with a known number of components; the paper itself shows that changing $\alpha$ and $k$ can turn two detected bursts into four, so parameter sensitivity is the first thing to check.
  • The same 'leaning as timestamp' trick could be applied to any bounded one-dimensional attitude distribution — vaccine sentiment, climate beliefs, party ratings — wherever binning choices and mode-distance sensitivity currently distort scalar measures.
  • If the burst count is interpreted as the number of distinct camps in a discussion, the paper's single-burst majority is a sharper version of the segregation claim: most comment sections are not two-sided battlegrounds but one-sided echo chambers, and the platform's global polarized look comes from averaging across sections.
  • Because only active users (at least five comments) enter the analysis, the findings describe the most engaged 2.1 million users; a natural extension is to check whether the mode structure survives when occasional commenters are included with a different weight.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The manuscript evaluates five polarization measures (Bimodality Coefficient, Hartigan's Dip Test, Distance from Unimodality, Van der Eijk's A, and Balance) on synthetic distributions and on a large corpus of YouTube comments from the 2020 U.S. presidential election period, and proposes an adaptation of Kleinberg's burst detection algorithm to count modes in distributions of inferred user leaning. The synthetic comparison shows that the Bimodality Coefficient can misclassify skewed unimodal distributions as bimodal, and the empirical application suggests that most comment sections are not strongly polarized under four of the five measures. The burst detection method is then applied to 65,928 comment sections, and the number of detected bursts is correlated with the polarization measures.

Significance. If the burst-detection adaptation were validated, it would provide a scalable tool for mode identification in opinion distributions, which would be a genuinely useful complement to aggregate polarization scores. The paper's comparative analysis of the five measures is transparent and internally consistent, and the large-scale YouTube case study is a valuable empirical contribution, particularly the demonstration that the Bimodality Coefficient's skewness sensitivity leads to qualitatively different conclusions. However, the central methodological claim is not yet supported: the method's mode-count accuracy is never tested against ground truth, and the parameters and preprocessing choices are not validated. The work is therefore significant only as a critical review plus a promising but unproven method.

major comments (5)
  1. [Supplementing the measures: Burst detection] The central claim that the adapted Kleinberg algorithm "improve[s] mode detection" is not tested against any ground truth. The paper reports no synthetic experiments with known numbers of modes; Figure 5 only illustrates parameter sensitivity, and Figure 6 correlates burst counts with the five polarization measures, which are themselves under test. Please add a validation experiment on distributions with known mode counts (e.g., mixtures of k Gaussians with k known), reporting mode-count accuracy and comparing against existing mode-detection baselines such as Hartigan's dip test or excess-mass bump hunting.
  2. [Supplementing the measures: Burst detection] The preprocessing step that rounds leaning scores to three decimals and adds epsilon = 0.0001 to non-unique scores creates inter-arrival times that are determined by the arbitrary tie-breaking constant. In dense regions with many duplicated scores, the pseudo-timestamps are spread over intervals whose width grows with the number of duplicates, so the burst structure may reflect the rounding/epsilon grid rather than the distribution's shape. Please demonstrate that the detected modes are invariant to epsilon (e.g., 1e-3, 1e-4, 1e-5) and to the number of rounding decimals, or replace this preprocessing with a principled weighted or jittered approach.
  3. [Supplementing the measures: Burst detection] The fixed parameters s = 1.7, gamma = 0.9, alpha = 0.05, and k = 0.5 are asserted without justification, and Figure 5a shows that changing alpha and k changes the number of detected bursts from two to four. The peak threshold of intensity >= 3 is imported from reference [20] without re-validation. Please provide a systematic sensitivity analysis and a principled parameter-selection procedure, for instance by optimizing or checking robustness on synthetic data with known modes.
  4. [Supplementing the measures: Burst detection] The claim that "the burst detection method aligns well with these measures" is based on Spearman correlations with the very measures the method is meant to complement. Since those measures are not ground truth for the number of modes, low-to-moderate correlations (rho between -0.36 and 0.26) do not validate mode detection. Please test the method against known mode counts or against independent mode-detection algorithms, and report the relationship to the measures only as a descriptive result.
  5. [Comparison on synthetic data] The comparative evaluation of the five measures uses only six hand-picked synthetic distributions with no repeated sampling or error bars, and the bin count K = 8 for DFU and Van der Eijk's A is chosen without justification. To support the claim that "no single measure is universally superior," please add a more systematic simulation that varies sample size, mode separation, skewness, and noise, and reports the measures' behavior over many replications.
minor comments (8)
  1. [Description of the selected measures] Equations (1) and (2) for the Bimodality Coefficient are typeset ambiguously: the standard formula is (skewness^2 + 1) / kurtosis, and the sample version in Eq. (2) needs parentheses and correct placement of the (n-1)^2/((n-2)(n-3)) term.
  2. [Description of the selected measures] In Eq. (3), the piecewise conditions for the difference values d_i appear inconsistent: "m < i < K" and "2 < i < m" should likely be "i > m" and "i < m", and the case i = m should set d_m = 0. Please clarify the indexing over i = 2, ..., K.
  3. [Case study on political conversations] The phrase "we are gonna utilize" is informal and should be replaced with "we will use".
  4. [Figure 3] In Figure 3 Panel B, the label "Comment section's leaning" and the legend ordering are confusing; also, the description of the DFU rescaling is repeated in the caption and the main text.
  5. [Supplementing the measures: Burst detection] The sentence "The burst-aggregating procedure are presented for two different levels of alpha and k" has subject-verb disagreement and should be revised.
  6. [Supplementing the measures: Burst detection] The peak threshold of intensity >= 3 is said to follow "previous studies [20]", but it is not explained why a threshold calibrated for burst detection in social media activity should transfer to mode detection in leaning distributions; please justify or re-derive this threshold.
  7. [Discussion and comments] The statement that "the burst detection method aligns well with these measures" overstates the reported correlations, which are described in the same section as "generally low to moderate"; please reconcile these two descriptions.
  8. [Methods] The paper does not include a data or code availability statement; for a computational study of this scale, a statement about availability of the scraping scripts, intermediate data, or analysis code is important for reproducibility.

Circularity Check

2 steps flagged · score 3.0 of 10

Burst-detection peak threshold is imported from an overlapping-author paper, and the method is validated against the very polarization measures it is meant to complement; no ground-truth mode-count test is given.

  1. ansatz smuggled in via citation [Supplementing the measures: Burst detection (peak definition after the aggregation step, parameters s=1.7, gamma=0.9, alpha=0.05, k=0.5)]
    "After completing the aggregation step, we define a peak, following previous studies [20], as any burst with an intensity level equal to or greater than 3."

    Reference [20] is Avalle et al., Nature 628, 582-589 (2024), whose author list includes M. Cinelli and W. Quattrociocchi, two of the present authors. The intensity threshold of 3 is the step that converts burst levels into the reported 'number of bursts/modes', which is the headline output of the proposed adaptation. The paper does not validate this threshold against known multimodal distributions; it simply inherits it from the self-authored prior study. Figure 5 shows that the number of detected bursts changes with the other aggregation parameters, and no experiment establishes that level-3 bursts correspond to true modes, so the central mode-count result rests on an unexamined, self-cited ansatz.

  2. other [Supplementing the measures: Burst detection (paragraph analyzing 65,928 comment sections and Figure 6)]
    "we observe a generally low to moderate correlation between burst counts and the scores of various polarization measures... This pattern suggests that the burst detection method aligns well with these measures."

    The burst method is introduced to supply shape information (number of modes) that the five measures lack, yet its success is judged by correlation with those same measures. Earlier in the same section the authors note that choosing among measures is 'far from being trivial, especially in absence of a ground truth'. The five measures are themselves under test and disagree with one another (e.g., Bimodality Coefficient vs. DFU), so agreement with them cannot establish that detected bursts are real modes. The correlation is empirical rather than definitionally forced, but using the very instruments the method was meant to complement as the benchmark is a self-referential validation, not an independent test.

full rationale

The paper's evaluation of the five standard polarization measures on synthetic and YouTube data is self-contained: those measures are computed from defined formulas and benchmarked against visually classified synthetic distributions, with no circular dependence on the new burst method. The proposed Kleinberg burst detection adaptation is also a genuinely new algorithmic contribution, with a toy example (Figure 4) showing that detected bursts coincide with visible modes. However, the central mode-count result is weakened by two self-referential elements. First, the definition of a peak as a burst of intensity at least 3 is imported via citation [20], a prior paper with overlapping authorship (Cinelli and Quattrociocchi), and is not re-validated against ground-truth mode counts; the number of detected modes is directly sensitive to this threshold and to the unvalidated alpha and k parameters. Second, the method is validated by correlating burst counts with the five polarization measures, even though the paper itself states there is no ground truth and the measures disagree with one another. These issues make the validation partly circular and the headline claim under-supported, but they do not reduce the derivation to a fit by construction, so the score is moderate rather than high.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The analysis rests on several assumptions: the MBFC channel labels are accurate; a user's average comment-channel leaning is a valid proxy for their ideology; rounding and jittering leaning scores does not distort the distribution; and bursts in the pseudo-time series correspond to modes. The method also depends on six hand-chosen parameters (s, gamma, alpha, k, peak threshold, and bin count). No independent ground truth is used to validate the mode detection.

free parameters (6)
  • s (Kleinberg smoothing parameter) = 1.7
    Chosen for the burst detection analysis; no sensitivity analysis or optimization against ground truth is reported.
  • gamma (burst deviation threshold) = 0.9
    Chosen for the burst detection analysis; controls the level of deviation from baseline required for a burst.
  • alpha (HDI coverage probability) = 0.05
    Used to compute the aggregation parameter phi; Fig. 5 shows different alpha values change the number of detected bursts.
  • k (aggregation multiplier) = 0.5
    Used to scale the HDI length into the aggregation threshold phi; Fig. 5 shows different k values change the number of detected bursts.
  • Peak intensity threshold = 3
    Bursts with intensity >= 3 are defined as peaks, following Avalle et al. (2024) [20]; this threshold is applied without re-validation on the present data.
  • Number of bins for DFU and Van der Eijk's A = 8
    The two binning-based measures are computed with eight bins for both synthetic and real data; no bin-count sensitivity analysis is provided.
assumptions (5)
  • domain assumption Channel political leaning labels from Media Bias/Fact Check via Wu and Resnick are accurate.
    Used to assign a leaning score to each comment; errors in these labels propagate to all user leaning estimates.
  • domain assumption A user's political leaning is the average of the leaning labels of the channels they comment on.
    This inference rule (Section 'Case study on political conversations') assumes commenting behavior reflects ideology and ignores comment content.
  • ad hoc to paper Rounding leaning scores to three decimals and adding epsilon does not materially distort the distribution.
    Introduced to make the pseudo-time series work with Kleinberg's algorithm; no analysis of distortion is given.
  • ad hoc to paper Bursts detected by Kleinberg's algorithm on the pseudo-time series correspond to modes of the underlying distribution.
    This is the core assumption of the proposed method, stated without proof or validation in Section 'Supplementing the measures: Burst detection'.
  • domain assumption Users with fewer than five comments can be excluded without biasing the analysis of polarization.
    The 'active users' threshold (at least five comments) is applied without a robustness check, even though the majority of users post fewer comments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying Polarization: A Comparative Study of Measures and Methods." pith.science (2026). https://pith.science/paper/UZGK4AGG

@misc{pith2026250107473,
  author       = {Pith},
  title        = {Pith review of: Quantifying Polarization: A Comparative Study of Measures and Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZGK4AGG}},
  note         = {Machine review of arXiv:2501.07473}
}
read the original abstract

Political polarization, a key driver of social fragmentation, has drawn increasing attention for its role in shaping online and offline discourse. Despite significant efforts, accurately measuring polarization within ideological distributions remains a challenge. This study evaluates five widely used polarization measures, testing their strengths and weaknesses with synthetic datasets and a real-world case study on YouTube discussions during the 2020 U.S. Presidential Election. Building on these findings, we present a novel adaptation of Kleinberg's burst detection algorithm to improve mode detection in polarized distributions. By offering both a critical review and an innovative methodological tool, this work advances the analysis of ideological patterns in social media discourse.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [20]

    Nature 628(8008), 582–589 (2024)

    Avalle, M., Di Marco, N., Etta, G., Sangiorgio, E., Alipour, S., Bonetti, A., Alvisi, L., Scala, A., Baronchelli, A., Cinelli, M., et al.: Persistent interaction patterns across social media platforms and over time. Nature 628(8008), 582–589 (2024)

    work page 2024

  2. [1]

    Accessed: 2024-10-03 (2024)

    World Economic Forum: Global Risks Report 2024. Accessed: 2024-10-03 (2024). https://www.weforum.org/publications/global-risks-report-2024/

    work page 2024

  3. [2]

    Annual review of political science 22(1), 129–146 (2019)

    Iyengar, S., Lelkes, Y., Levendusky, M., Malhotra, N., Westwood, S.J.: The origins and consequences of affective polarization in the united states. Annual review of political science 22(1), 129–146 (2019)

    work page 2019

  4. [3]

    Public opinion quarterly 76(3), 405–431 (2012)

    Iyengar, S., Sood, G., Lelkes, Y.: Affect, not ideology: A social identity perspective on polarization. Public opinion quarterly 76(3), 405–431 (2012)

    work page 2012

  5. [4]

    i disrespectfully agree

    Mason, L.: “i disrespectfully agree”: The differential effects of partisan sorting on social and issue polarization. American journal of political science 59(1), 128–145 (2015)

    work page 2015

  6. [5]

    Ann Rev Polit Sci (2014)

    Center, P.R.: Political polarization in the american public. Ann Rev Polit Sci (2014)

    work page 2014

  7. [6]

    Nature Communications 15(1), 9560 (2024)

    Falkenberg, M., Zollo, F., Quattrociocchi, W., Pfeffer, J., Baronchelli, A.: Pat- terns of partisan toxicity and engagement reveal the common structure of online political communication across countries. Nature Communications 15(1), 9560 (2024)

    work page 2024

  8. [7]

    Nature Climate Change 12(12), 1114–1121 (2022)

    Falkenberg, M., Galeazzi, A., Torricelli, M., Di Marco, N., Larosa, F., Sas, M., Mekacher, A., Pearce, W., Zollo, F., Quattrociocchi, W., et al.: Growing polar- ization around climate change on social media. Nature Climate Change 12(12), 1114–1121 (2022)

    work page 2022

Show all 45 references
  1. [8]

    Communications of the ACM 47(12), 57–59 (2004)

    Sunstein, C.R.: Democracy and filtering. Communications of the ACM 47(12), 57–59 (2004)

    2004

  2. [9]

    The Wall Street Journal (2021)

    Wall Street Journal Staff: The facebook files. The Wall Street Journal (2021). Accessed: 2024-11-08

    2021

  3. [10]

    Proceedings of the National Academy of Sciences 118(9), 2023301118 (2021)

    Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quattrociocchi, W., Starnini, M.: The echo chamber effect on social media. Proceedings of the National Academy of Sciences 118(9), 2023301118 (2021)

    2021

  4. [11]

    Current Opinion in Psychology 47, 101407 (2022)

    Cinelli, M., Etta, G., Avalle, M., Quattrociocchi, A., Di Marco, N., Valensise, C., Galeazzi, A., Quattrociocchi, W.: Conspiracy theories and social media platforms. Current Opinion in Psychology 47, 101407 (2022)

    2022

  5. [12]

    arXiv preprint arXiv:2411.16826 (2024)

    Di Martino, E., Galeazzi, A., Starnini, M., Quattrociocchi, W., Cinelli, M.: Characterizing the fragmentation of the social media ecosystem. arXiv preprint arXiv:2411.16826 (2024)

    2024 arXiv

  6. [13]

    Barber´ a, P., Jost, J.T., Nagler, J., Tucker, J.A., Bonneau, R.: Tweeting from 18 left to right: Is online political communication more than an echo chamber? Psychological science 26(10), 1531–1542 (2015)

    2015

  7. [14]

    Contemporary Climate Change Debates: a student primer

    North, P., Schaefer, M.: Are social media making constructive climate pol- icy making harder. Contemporary Climate Change Debates: a student primer. Cambridge, UK: Polity (2019)

    2019

  8. [15]

    Philosophy of science 84(1), 115–159 (2017)

    Bramson, A., Grim, P., Singer, D.J., Berger, W.J., Sack, G., Fisher, S., Flocken, C., Holman, B.: Understanding polarization: Meanings, measures, and model evaluation. Philosophy of science 84(1), 115–159 (2017)

    2017

  9. [16]

    Information Sciences 588, 265–278 (2022)

    Arruda, H.F., Cardoso, F.M., Arruda, G.F., Hern´ andez, A.R., Fontoura Costa, L., Moreno, Y.: Modelling how social network algorithms can influence opinion polarization. Information Sciences 588, 265–278 (2022)

    2022

  10. [17]

    Journal of the Korean Physical Society 80(6), 509–515 (2022)

    Kim, J., Baek, S.K.: Democracy and polarization in the national assembly of the republic of korea. Journal of the Korean Physical Society 80(6), 509–515 (2022)

    2022

  11. [18]

    Cognitive Computation 15(2), 731–738 (2023)

    Pavlopoulos, J., Likas, A.: Distance from unimodality for the assessment of opinion polarization. Cognitive Computation 15(2), 731–738 (2023)

    2023

  12. [19]

    Proceedings of the ACM on Human- Computer Interaction 4(CSCW1), 1–27 (2020)

    Hussein, E., Juneja, P., Mitra, T.: Measuring misinformation in video search platforms: An audit study on youtube. Proceedings of the ACM on Human- Computer Interaction 4(CSCW1), 1–27 (2020)

    2020

  13. [21]

    Institute, S.: SAS/STAT User’s Guide: GLM-V ARCOMP vol. 2. SAS institute Incorporated, ??? (1990)

    1990

  14. [22]

    American Journal of Botany 74(8), 1280–1288 (1987)

    Ellison, A.M.: Effect of seed dimorphism on the density-dependent dynamics of experimental populations of atriplex triangularis (chenopodiaceae). American Journal of Botany 74(8), 1280–1288 (1987)

    1987

  15. [23]

    Frontiers in psychology 4, 700 (2013)

    Pfister, R., Schwarz, K.A., Janczyk, M., Dale, R., Freeman, J.B.: Good things peak in pairs: a note on the bimodality coefficient. Frontiers in psychology 4, 700 (2013)

    2013

  16. [24]

    Journal of Modern Applied Statistical Meth- ods 6(1), 3 (2007)

    Knapp, T.R.: Bimodality revisited. Journal of Modern Applied Statistical Meth- ods 6(1), 3 (2007)

    2007

  17. [25]

    The annals of Statistics, 70–84 (1985)

    Hartigan, J.A., Hartigan, P.M.: The dip test of unimodality. The annals of Statistics, 70–84 (1985)

    1985

  18. [26]

    Behavior research methods 45, 83–97 (2013) 19

    Freeman, J.B., Dale, R.: Assessing bimodality to detect the presence of a dual cognitive process. Behavior research methods 45, 83–97 (2013) 19

    2013

  19. [27]

    Quality and Quantity 35, 325–341 (2001)

    Eijk, C.: Measuring agreement in ordered rating scales. Quality and Quantity 35, 325–341 (2001)

    2001

  20. [28]

    Vaccine 36(25), 3606–3612 (2018)

    Schmidt, A.L., Zollo, F., Scala, A., Betsch, C., Quattrociocchi, W.: Polarization of the vaccination debate on facebook. Vaccine 36(25), 3606–3612 (2018)

    2018

  21. [29]

    Stocking, G., Van Kessel, P., Barthel, M., Matsa, K.E., Khuzam, M.: Many amer- icans get news on youtube, where news organizations and independent producers thrive side by side (2020)

    2020

  22. [30]

    In: Proceedings of the Inter- national AAAI Conference on Web and Social Media, vol

    Wu, S., Resnick, P.: Cross-partisan discussions on youtube: Conservatives talk to liberals but liberals don’t talk to conservatives. In: Proceedings of the Inter- national AAAI Conference on Web and Social Media, vol. 15, pp. 808–819 (2021)

    2021

  23. [31]

    In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp

    Ribeiro, M.H., Ottoni, R., West, R., Almeida, V.A., Meira Jr, W.: Auditing radicalization pathways on youtube. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp. 131–141 (2020)

    2020

  24. [32]

    arXiv e-prints, 2308 (2023)

    Hammas Saeed, M., Papadamou, K., Blackburn, J., De Cristofaro, E., Stringhini, G.: Tuberaider: Attributing coordinated hate attacks on youtube videos to their source communities. arXiv e-prints, 2308 (2023)

    2023

  25. [33]

    Proceedings of the ACM on Human-Computer Interaction 5(CSCW2), 1–25 (2021)

    Papadamou, K., Zannettou, S., Blackburn, J., De Cristofaro, E., Stringhini, G., Sirivianos, M.: ” how over is it?” understanding the incel community on youtube. Proceedings of the ACM on Human-Computer Interaction 5(CSCW2), 1–25 (2021)

    2021

  26. [34]

    Political analysis 23(1), 76–91 (2015)

    Barber´ a, P.: Birds of the same feather tweet together: Bayesian ideal point estimation using twitter data. Political analysis 23(1), 76–91 (2015)

    2015

  27. [35]

    In: Proceedings of the International AAAI Conference on Web and Social Media, vol

    Cossard, A., Morales, G.D.F., Kalimeri, K., Mejova, Y., Paolotti, D., Starnini, M.: Falling into the echo chamber: The italian vaccination debate on twitter. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 14, pp. 130–140 (2020)

    2020

  28. [36]

    In: Proceedings of the International AAAI Conference on Web and Social Media, vol

    Zhou, D.X., Resnick, P., Mei, Q.: Classifying the political leaning of news articles and users from user votes. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 5, pp. 417–424 (2011)

    2011

  29. [37]

    In: Proceedings of the International AAAI Conference on Web and Social Media, vol

    Hemphill, L., Sch¨ opke-Gonzalez, A.M.: Two computational models for analyzing political attention in social media. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 14, pp. 260–271 (2020)

    2020

  30. [38]

    In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (volume 1: Long 20 Papers), pp

    Preot ¸iuc-Pietro, D., Liu, Y., Hopkins, D., Ungar, L.: Beyond binary labels: Polit- ical ideology prediction of twitter users. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (volume 1: Long 20 Papers), pp. 729–740 (2017)

    2017

  31. [39]

    Journal of communi- cation 60(3), 556–576 (2010)

    Stroud, N.J.: Polarization and partisan selective exposure. Journal of communi- cation 60(3), 556–576 (2010)

    2010

  32. [40]

    The american statistician 27(1), 17–21 (1973)

    Anscombe, F.J.: Graphs in statistical analysis. The american statistician 27(1), 17–21 (1973)

    1973

  33. [41]

    In: Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp

    Kleinberg, J.: Bursty and hierarchical structure in streams. In: Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 91–101 (2002)

    2002

  34. [42]

    Nature communications 10(1), 7 (2019)

    Bovet, A., Makse, H.A.: Influence of fake news in twitter during the 2016 us presidential election. Nature communications 10(1), 7 (2019)

    2019

  35. [43]

    In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp

    Stefanov, P., Darwish, K., Atanasov, A., Nakov, P.: Predicting the topical stance and political leaning of media using tweets. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 527–537 (2020)

    2020

  36. [44]

    Lewis, R.: Alternative influence: Broadcasting the reactionary right on youtube (2018)

    2018

  37. [45]

    arXiv preprint arXiv:1912.11211 (2019) 21

    Ledwich, M., Zaitsev, A.: Algorithmic extremism: Examining youtube’s rabbit hole of radicalization. arXiv preprint arXiv:1912.11211 (2019) 21

    2019 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.