Pith. sign in

REVIEW 4 major objections 8 minor 23 references

Using Social Media for Word-of-Mouth Marketing

T0 review · 4 major / 8 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read A topic-sensitive PageRank on a reaction-weighted social interaction graph identifies the most influential Facebook group members for word-of-mouth marketing, beating HITS, Z-score, eigenvector, betweenness, and closeness.

desk verdict Plausible but under-validated application of PageRank to Facebook groups; the main evidence is partly circular and the independent check is thin, but the dataset and descriptive analysis give it enough substance to deserve a conditional referee. read the letter →

arxiv 1908.08298 v1 pith:XWQUTO4O submitted 2019-08-22 cs.SI cs.IR

classification cs.SIcs.IR
keywords word-of-mouthmarketinginfluentialusersPageRankFacebookgroupstopic-sensitiverankingsocialinteractiongraphreinforcedcentralitycomparison
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

Marketing on social media often fails because users distrust ads; word-of-mouth from a trusted peer works better. This paper argues that in focused Facebook groups, the right people to recruit for word-of-mouth marketing are the authorities who already attract most reactions, and that a topic-sensitive PageRank on a graph of user reactions finds them reliably. The authors build a social interaction graph whose edge weights combine the type of reaction (like on comment, like, comment, share) with how semantically relevant the post is to the advertised topic, then rank members by PageRank. On 100 groups with 0.3 million posts and 10 million reactions, this outranks HITS, Z-score, eigenvector, betweenness, and closeness in matching human judgments (MAP 0.91, NDCG 0.83). They also propose reinforced marketing, where several authorities in the same subgroup promote a product together, and identify the best months to launch.

What carries the argument

The central object is the topic-sensitive social interaction graph: vertices are group members, and each directed edge from one user to another records that the first user reacted to content created by the second. Edge weight is the product of a reaction-type weight (1, 2, 4, 8 for like on comment, like, comment, share) and a boosted relevance that grows logarithmically with the summed mutual-information similarity between topic words and words in the post. PageRank runs on this weighted graph and propagates authority through the reaction structure, so a user ranks highly not merely by receiving many reactions but by receiving reactions from users who are themselves highly ranked. The same graph supplies sub-group detection via weakly connected components for reinforced marketing.

What would settle it

Give matched groups of Facebook members the same product recommendation from a top-PageRank user and from a randomly selected low-rank user and measure actual sign-ups or purchases; if the low-rank user converts as well as the top-ranked one, the claim that PageRank identifies influential marketers for word-of-mouth is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that PageRank, applied to a topic-sensitive social interaction graph whose edges are weighted by reaction type and topical relevance, identifies the most influential users for word-of-mouth marketing in question-answer style Facebook groups. The authors treat a reaction from user A to user B's post or comment as an edge from A to B with weight equal to the reaction's assigned importance (1 for like on comment, 2 for like, 4 for comment, 8 for share) multiplied by a relevance boost derived from semantic similarity between the post and the advertised topic. Ranking group members by PageRank on this graph yields the top authorities, and the paper reports that this ranking agrees with human expert judgments better than HITS, Z-score, eigenvector centrality, betweenness, and closeness (MAP 0.91, NDCG 0.83). The same framework supports reinforced marketing, in which each selected authority is backed by at least r minus 1 other authorities in the same weakly connected subgroup, so repeated trusted messages reinforce the product.

Load-bearing premise

The framework assumes that the reactions a user receives—likes, comments, shares, and likes on comments—measure that user's real ability to persuade others to buy, because both the PageRank graph and the votes used to validate it are built from the same reactions.

Editorial extensions

If this is right

  • An advertiser can rank the top-k potential promoters in a large topic-focused Facebook group from public reaction data alone, without surveying members.
  • Reinforced marketing gives a selection rule: for each weakly connected subgroup that is large enough, pick its top-r authorities so that every chosen promoter is backed by at least r-1 other trusted voices within that subgroup.
  • Because top users rank highly across related topics, the same authorities can promote a range of products in their group, not just a single topic.
  • Ranking quality degrades when the advertised topic is weakly related to the group's shared interest, so marketers should choose groups whose core topic matches the product.
  • Posting and reaction peaks in March, April, and October in technical groups suggest those months are the best launch windows, at least in the examined population.

Reading between the lines

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

  • A testable extension: the same graph construction could be applied to other question-answer communities where reactions are observable, and the ranking validated against sign-ups or purchases rather than reactions.
  • The fixed reaction weights (1,2,4,8) could be treated as free parameters and fit to outcome data; the ranking's sensitivity to those weights is not examined in the paper.
  • If the mechanism generalizes, marketers could combine reinforced marketing with temporal analysis to schedule launches in periods when top users are already most active.
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.

Referee Report

4 major / 8 minor

Summary. The paper proposes a method for word-of-mouth marketing in Facebook groups. It constructs a topic-sensitive social interaction graph from user posts and reactions, assigns edge weights based on reaction type and boosted topical relevance, and ranks users with PageRank. The method is compared against HITS, Z-score, eigenvector centrality, betweenness, and closeness using Pearson correlation with votes and topical votes, and with MAP/NDCG against manual labels from five students. The paper reports that PageRank outperforms the other measures, and also sketches a reinforced marketing strategy based on weakly connected components, analyzes influence across topics and time, and recommends marketing periods. The main claimed contribution is that PageRank identifies influential users suitable for word-of-mouth marketing in online social groups.

Significance. If the central claim holds, the paper offers a practical and scalable way to identify potential word-of-mouth marketers in focused Facebook groups using only publicly observable reactions, which would be useful for marketing practitioners and social-media analysts. Strengths include the large dataset (100 groups, 0.3 million posts, 10 million reactions), a clearly described topic-sensitivity mechanism, and a systematic comparison across six ranking algorithms. The manual-label evaluation in Section 8.4 is genuinely independent of the graph construction and gives some support for the plausibility of PageRank. However, as detailed below, the primary validation evidence is partly circular because the same reaction data is used to build the graph and to define the popularity baselines, and the external validation is too small and underspecified to fully carry the word-of-mouth effectiveness claim. The paper is best read as evidence about predicting reaction-weighted popularity rather than about persuasion or purchase influence.

major comments (4)
  1. [§5.1 vs. §8.2–8.3] The main evaluation is partly circular. In Section 5.1, graph edge weights are computed as the product of a reaction-type weight (like on comment=1, like=2, comment=4, share=8) and the boosted topical relevance. In Section 8.2, the baseline 'votes' and 'topical votes' are weighted sums of the same reaction types over the same posts and comments, using the same reaction weights. Thus votes are a near-linear function of a user's weighted indegree, and PageRank is computed on a graph whose edge weights are built from the same quantities. The high correlations in Figures 4 and 5 therefore partly measure self-consistency rather than external marketing influence. To support the claim, the authors should compare against a direct weighted-indegree baseline or total-votes baseline, and should clearly state that the evidence concerns reaction-weighted popularity, not persuasion or purchase behavior.
  2. [§8.4] The only non-circular validation is the manual labeling experiment, but it is too thin to carry the central claim. Five students from the authors' own research lab judged whether users are influential, and they were given the posts and reactions of those users. The labels can therefore inherit the same visibility and popularity bias that the reaction-based metrics already encode, and they measure perceived influence rather than actual word-of-mouth effectiveness. No error bars, confidence intervals, or significance tests are reported for the MAP/NDCG values in Table 1, and the 92% agreement figure is not accompanied by a chance-adjusted measure such as Cohen's kappa. The authors should report per-group variability, inter-annotator agreement statistics, and a precise definition of the judgment task.
  3. [§7] The reinforced marketing contribution is described but never empirically evaluated. The algorithm depends on the subgroup threshold th and the reinforcement parameter r, but no values, experiments, or sensitivity analyses are provided, and no results show that selecting r top users per weakly connected component improves influence or marketing outcomes. Since reinforced marketing is presented as a key contribution in the abstract and in Problem 3, the authors should either provide an evaluation or clearly state that this part is a proposal rather than a validated result.
  4. [§5.1, §6, §8.1] The method has several free parameters whose robustness is not examined: the boost factor alpha (set to 20), the reaction-type weights 1, 2, 4, 8, the PageRank damping factor and iteration count, and the subgroup threshold th. The reported superiority of PageRank in Table 1 and Figures 4–5 could depend on these choices. A sensitivity analysis, or at least a justification for the chosen values, is needed before the claim that PageRank is generally the best authority measure for this task can be accepted.
minor comments (8)
  1. [Abstract and §8.1] The abstract says '0.3 million messages' while Section 8.1 says '0.3 million posts'; please make the terminology consistent.
  2. [§8.3] The description says Pearson correlation is used on ranks, but Pearson is not a rank correlation unless explicitly applied to ranks; please clarify whether raw scores or rank-transformed scores were correlated, and consider reporting Spearman correlation.
  3. [§8.2–8.3] Figures 4 and 5 lack axis labels, error bars, and information about the number of groups used for averaging; adding these would make the comparisons interpretable.
  4. [§5.1] The Semantic Link system is cited only by a URL; please provide a version, a description of its training data, and a statement about how mutual-information scores are normalized across words and posts.
  5. [§8.4] The manual labeling procedure is described too briefly: the paper should state how the students were instructed to define 'influential', how many users each student labeled per group, and whether the labels were collected independently of the ranking algorithms.
  6. [§8.5] The correlations in Table 2 are reported without sample sizes or significance levels, so the claim that high top-20 correlation indicates multi-topic influence is not statistically supported.
  7. [§2 and §8.6] The statistics '6.5% users generate 85% of content' and 'less than 2% influence 80% of the population' are asserted without a clear definition of 'influence' or a description of how the 80% figure is computed; please clarify.
  8. [General] No data or code availability statement is provided. Releasing the dataset (at least in anonymized form) and the ranking implementation would substantially improve reproducibility and allow independent verification of the reported results.

Circularity Check

1 steps flagged · score 6.0 of 10

Evaluation is partly circular: the §5 graph edges and the §8.2 'votes'/'topical votes' baselines are weighted sums of the same reaction counts, so PageRank's reported correlation mostly measures self-consistency; the §8.4 manual labels are the only external check.

  1. self definitional [Section 5.1 (graph construction) and Section 8.2 (influence baselines)]
    "We assign the weights 1, 2, 4 and 8 for like on comment, like, comment, and share respectively [2]. ... For each user, we compute votes by taking the weighted sum of all the audience reactions received by the user over all his posts, comments. However, we compute topical votes by taking the weighted sum of audience reactions over all his posts, comments that contain the advertisement topic itself or the topics semantically related to the advertisement topic."

    The edge weights of the topic-sensitive graph are defined from exactly the same reaction counts (weighted by reaction type and boosted relevance) that make up the validation baselines: 'votes' is the weighted sum of reactions received, and 'topical votes' restricts that same weighted sum to topic-related posts. PageRank is then correlated against this baseline in Figures 4 and 5, so the reported agreement largely measures whether PageRank reproduces a non-recursive aggregate of its own input edge weights. It does not provide independent evidence that the ranked users can persuade anyone or generate marketing outcomes; the only external check is the manual student labeling in Section 8.4.

full rationale

The central derivation (PageRank on a topic-sensitive reaction-weighted graph) is not mathematically circular in itself: PageRank is a standard algorithm, and the graph weights are stated independently of the validation baselines. The circularity enters at the evaluation stage. Section 5.1 defines edge weights as reaction-type weights (1, 2, 4, 8) multiplied by boosted topical relevance; Section 8.2 defines the 'votes' and 'topical votes' baselines as weighted sums of the same audience reactions (all posts or topic-related posts). Consequently, the strong correlations claimed in Figures 4 and 5 are largely a test of whether PageRank's ranking reproduces a local weighted-sum of its own edge weights, not an independent test of word-of-mouth influence. The manual student labels in Section 8.4 are a genuinely external check and give the best evidence (MAP 0.91 / NDCG 0.83), but the labels are based on the same posts and reactions and measure perceived influence, not persuasion or purchase behavior. No load-bearing self-citation occurs. The result is thus partially circular rather than fully degenerate, supporting a score of 6.

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

The paper's graph construction, ranking, and validation all rest on the assumption that Facebook reactions encode influence and on an external semantic-relatedness service; these are not independently verified. The free parameters are chosen by hand without sensitivity analysis.

free parameters (4)
  • alpha (α) in boosted relevance = 20
    Set by hand in Algorithm 1 line 8; scales the logarithmic boost bRelevance = 1 + α*ln(1 + relevance), with no sensitivity analysis.
  • Reaction-type weights = like on comment: 1, like: 2, comment: 4, share: 8
    Assigned in Section 5 following citation [2]; these weights determine all edge weights and PageRank results but are not learned or justified for this dataset.
  • Subgroup threshold th = not specified
    Section 7 selects weakly connected components for reinforced marketing only if they contain enough users, but the threshold value is never given.
  • PageRank damping factor and iteration count = not specified
    Standard PageRank parameters are not reported; rankings can change with the damping factor.
assumptions (4)
  • domain assumption Reactions (likes, comments, shares) are a valid proxy for trust and marketing influence.
    Section 5 builds graph edges only from reactions, and Section 8.2 uses weighted reaction sums as the ground-truth 'votes'. The paper never validates that reactions predict purchasing or promotional behavior.
  • domain assumption Semantic Link's Wikipedia-based mutual information accurately measures topical relevance of posts.
    Section 5.1 uses this external system to compute bRelevance; no accuracy check, version, or parameter details are provided.
  • domain assumption Manual judgments by five students are a reliable ground truth for influence.
    Section 8.4 uses these labels to compute MAP/NDCG after students were shown the same posts and reactions, which can bias labels; only average agreement is reported.
  • domain assumption Weakly connected components correspond to meaningful marketing subgroups.
    Section 7 assumes each component is a coherent audience segment for reinforced marketing, but component structure is not validated against actual product adoption or trust relationships.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Using Social Media for Word-of-Mouth Marketing." pith.science (2026). https://pith.science/paper/XWQUTO4O

@misc{pith2026190808298,
  author       = {Pith},
  title        = {Pith review of: Using Social Media for Word-of-Mouth Marketing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XWQUTO4O}},
  note         = {Machine review of arXiv:1908.08298}
}
read the original abstract

Nowadays online social networks are used extensively for personal and commercial purposes. This widespread popularity makes them an ideal platform for advertisements. Social media can be used for both direct and word-of-mouth (WoM) marketing. Although WoM marketing is considered more effective and it requires less advertisement cost, it is currently being under-utilized. To do WoM marketing, we need to identify a set of people who can use their authoritative position in social network to promote a given product. In this paper, we show how to do WoM marketing in Facebook group, which is a question answer type of social network. We also present concept of reinforced WoM marketing, where multiple authorities can together promote a product to increase the effectiveness of marketing. We perform our experiments on Facebook group dataset consisting of 0.3 million messages and 10 million user reactions.

Figures

Figures reproduced from arXiv: 1908.08298 by the authors.

Figure 2
Figure 2. Degree distribution in groups [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Social Interaction Graph users and edges as reactions. Weights of each type of reaction are assigned based on type of interaction as described in Section 5. We do not consider brelevance in this example for ease of understanding. We apply PageRank on unweighted version of this graph, we get rank of each node as x1 = 0.7210, x2 = 0.2403, x3 = 0.5408, x4 = 0.3605, whereas we get x1=0.7328, x2=0.1466, x3=0.5374, x4=0.3… view at source ↗
Figure 4
Figure 4. Correlation of authority finding algorithms with votes [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Correlation of top users across variety of topics We observe in [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Structure of Java group [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Posting behavior of top users [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    Computer networks (2000)

    Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Stata, R., Tomkins, A., Wiener, J.: Graph structure in the web. Computer networks (2000)

  2. [2]

    new media & society 14(7) (2012)

    Bucher, T.: Want to be on the top? algorithmic power and the threat of invisibility on facebook. new media & society 14(7) (2012)

  3. [3]

    In: SIGKDD

    Chen, W., Wang, Y., Yang, S.: Efficient influence maximization in social networks. In: SIGKDD. ACM (2009)

  4. [4]

    ACM (2014)

    Cheng, J., Adamic, L., Dow, P.A., Kleinberg, J.M., Leskovec, J.: Can cascades be predicted? In: WWW. ACM (2014)

  5. [5]

    In: SIGKDD

    Domingos, P., Richardson, M.: Mining the network value of customers. In: SIGKDD. ACM (2001)

  6. [6]

    Forbes: What are they saying about your brand?http://www.forbes.com/sites/ pauljankowski/2013/02/27/quick-what-are-they-saying-about-your-brand/ #ee5ff7374a8d (2013)

  7. [7]

    Social net- works (1978)

    Freeman, C, L.: Centrality in social networks conceptual clarification. Social net- works (1978)

  8. [8]

    Sociometry (1977)

    Freeman, L.C.: A set of measures of centrality based on betweenness. Sociometry (1977)

Show all 23 references
  1. [9]

    SIGMOD 42(2) (2013)

    Guille, A., Hacid, H., Favre, C., Zighed, D.A.: Information diffusion in online social networks: A survey. SIGMOD 42(2) (2013)

  2. [10]

    http://www.incite-group.com/brand-management/ how-social-media-amplifies-power-word-mouth (2014)

    Incite: How social media amplifies the power of word- of-mouth. http://www.incite-group.com/brand-management/ how-social-media-amplifies-power-word-mouth (2014)

  3. [11]

    ACM Transactions on Information Systems (TOIS) 20(4) (2002)

    Järvelin, K., Kekäläinen, J.: Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS) 20(4) (2002)

  4. [12]

    In: SIGKDD

    Kempe, D., Kleinberg, J., Tardos, É.: Maximizing the spread of influence through a social network. In: SIGKDD. ACM (2003)

  5. [13]

    In: International Computing and Combinatorics Conference

    Kleinberg, J.M., Kumar, R., Raghavan, P., Rajagopalan, S., Tomkins, A.S.: The web as a graph: measurements, models, and methods. In: International Computing and Combinatorics Conference. Springer (1999)

  6. [14]

    Biometrics (1989)

    Lawrence, I., Lin, K.: A concordance correlation coefficient to evaluate repro- ducibility. Biometrics (1989)

  7. [15]

    Leskovec, J., McGlohon, M., Faloutsos, C., Glance, N.S., Hurst, M.: Patterns of cascading behavior in large blog graphs. In: SDM. SIAM (2007)

  8. [16]

    Page, L., Brin, S., Motwani, R., Winograd, T.: The pagerank citation ranking: bringing order to the web. (1999)

  9. [17]

    Ruhnau, B.: Eigenvector-centrality—a node-centrality? Social networks (2000)

  10. [18]

    Journal of Marketing Research 47(4) (2010)

    Trusov,M.,Bodapati,A.V.,Bucklin,R.E.:Determininginfluentialusersininternet social networks. Journal of Marketing Research 47(4) (2010)

  11. [19]

    In: Semantic Hyper/Multimedia Adaptation

    Vogiatzis, D.: Influential users in social networks. In: Semantic Hyper/Multimedia Adaptation. Springer (2013)

  12. [20]

    In: Data Mining

    Wang, X., McCallum, A., Wei, X.: Topical n-grams: Phrase and topic discovery, with an application to information retrieval. In: Data Mining. ICDM 2007

  13. [21]

    In: WSDM

    Weng, J., Lim, E.P., Jiang, J., He, Q.: Twitterrank: finding topic-sensitive influen- tial twitterers. In: WSDM. ACM (2010)

  14. [22]

    Wu, S., Hofman, J.M., Mason, W.A., Watts, D.J.: Who says what to whom on twitter. In: WWW. ACM (2011)

  15. [23]

    Zhang, J., Ackerman, M.S., Adamic, L.: Expertise networks in online communities: structure and algorithms. In: WWW. ACM (2007)

Pith tools

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