Ranking users, papers and authors in online scientific communities

• The growth and specialization of science make it difficult for researchers to follow all relevant developments.Source: Introduction. – · p. 1
• Online scientific communities facilitate information exchange but do not eliminate the challenge of finding relevant papers and authors.Source: Introduction. – · p. 1
• Reputation systems compute entity scores to reduce information asymmetry and can create incentives for good behavior.Source: Introduction. – · p. 1
• The proposed algorithm simplifies QTR by omitting user social networks, then adds author credit computed jointly with paper quality and user reputation.Source: Introduction. – · p. 1
• Using citation count, impact factor, and h-index, the authors report that the extended algorithm outperforms other state-of-the-art algorithms.Source: Introduction. – · p. 1

• The community is modeled as a weighted bipartite user-item network, where link weights encode interaction importance or intensity.Source: Algorithms. – · pp. 1–2
• biHITS assigns reputation to users and quality to items through mutually dependent scores solved iteratively.Source: Algorithms. – · p. 2
• The iterative solution is normalized and stopped when total absolute changes fall below 10^-8; connected graphs yield a unique solution independent of initialization.Source: Algorithms. – · p. 2
• Weighted interactions can replace binary links, allowing the algorithm to incorporate interaction intensity.Source: Algorithms. – · p. 2
• QR simplifies QTR by omitting trust among users while retaining joint estimation of item quality and user reputation.Source: Algorithms. – · p. 2
• The parameters control sum-versus-average aggregation and whether interactions with inferior-quality items reduce reputation.Source: Algorithms. – · p. 2
• HITS-like methods require accumulated feedback and lack mechanisms to correct attention attracted for quality-unrelated reasons.Source: Algorithms with author credit · p. 2

Algorithms with author credit.

• HITS-like methods can rank attention attracted by witty or provocative titles despite lacking mechanisms to correct for quality-unrelated attention.Source: Algorithms with author credit · p. 2
• EigenRumor addresses this issue by adding information-provider scores, adapted here as author credit for communities containing users, papers, and authors.Source: Algorithms with author credit · p. 2
• The author-paper network uses matrix P, whose binary links indicate authorship or co-authorship between author m and paper α.Source: Algorithms with author credit · p. 2
• EigenRumor combines user reputation, author credit, and paper quality, with ω determining the relative contribution of authors and users to quality.Source: Algorithms with author credit · p. 2
• QRC extends QR with author credit, using λ analogously to ω; when λ = 0, quality and reputation remain QR values, while λ > 0 makes all three scores interdependent.Source: Algorithms with author credit · p. 2
• Unlike EigenRumor, QRC uses three scores and separate normalization exponents for user actions, authored papers, and co-authors.Source: Algorithms with author credit · pp. 2–3

Model construction

• The agent-based benchmark assigns users abilities and activity levels, items intrinsic fitness, and preferential links between able users and high-fitness items.These latent values provide the true quantities against which estimated reputation and quality are evaluated.Source: Model evaluation on artificial data. – · p. 3
• The simulated system evolves through user activity, item uploads, downloads, and fitness-dependent item selection, with fixed users and growing items and links.Its expected growth is M(t) = N⟨ν⟩pUt, while network density remains constant.Source: Model evaluation on artificial data. – · p. 3

Simulation settings

• The simulations use µ = 1/2, X = 1/2, h = 5, N = 1000, pU = 0.1, and t = 200, producing approximately 6,700 items and network density η ≈2%.Uploads receive weight 1 and downloads weight 0.1, reflecting greater importance assigned to uploading.Source: Model evaluation on artificial data. – · p. 3

Evaluation metrics

• The benchmark measures correlations with true quality and ability, while correlations with time and activity quantify bias; high values are desirable for the former and near-zero values for the latter.The four metrics are cQf, cRa, cQt, and cRν.Source: Model evaluation on artificial data. – · p. 3

Benchmark comparison

• In the selected comparison, biHITS correlates least with user ability and item quality while showing strong bias toward old items and especially active users.The result indicates that heterogeneous age and activity require additional normalization beyond the basic biHITS formulation.Source: Results on artificial data · p. 3, Table 1 · p. 3

Parameter selection

• QR2 outperforms QR1 in all four correlation metrics because ρQ = 1 penalizes users for uploading or downloading inferior content.Both settings share θQ = 0 and ρR = 0, while θR = 1 reduces reputation bias toward user activity.Source: Results on artificial data · p. 3
• The QR evaluation covers all 16 combinations of four binary parameters, excluding one numerically unstable configuration with no reportable results.The best-performing settings share θQ = 0 and ρR = 0.Source: Results on artificial data · p. 3

Limitations

• QR1 and QR2 both perform badly on cQt, showing strongly negative correlation between estimated quality and item age.The study suggests this age bias is likely to improve when real systems impose eventual saturation on item-degree growth.Source: Results on artificial data · p. 4

Parameter influence

• Performance improves with h and saturates near h ≃5, whereas µ is crucial because too few able or active users prevent reliable detection of quality content.Network sparsity is harmful when too low or too high, and the download-to-upload weight ratio ξ is nearly optimal at 0.1.Source: Results on artificial data · p. 4

Real-data evaluation

• The real-data evaluation uses Econophysics Forum logs covering 1,000 days from 6 July 2010 through 31 March 2013.The forum serves interdisciplinary physics researchers and finance specialists.Source: Model evaluation on real data. – · p. 4, Data description and analysis · p. 4

Data processing

• The analysis retains uploads, downloads, and abstract views while removing web-bot traffic, incomplete older records, and minimally active users.For repeated user-paper interactions, only the earliest interaction is retained because it is treated as the best indicator of initial interest.Source: Data description and analysis · p. 4

Data analysis

• Figure 2 presents cumulative degree distributions for users, papers, and authors across different interaction types, excluding the editor from the user-upload distribution for clarity.The figure provides the distributional overview used to characterize activity and attention in the forum.Source: Figure 2 · p. 4

Data description

• The final dataset contains 5,071 users, 844 papers, 29,748 links, and 1,527 normalized author identities, with citation counts and journal impact factors added for evaluation.The editor uploaded 85% of the papers in the analyzed sample.Source: Data description and analysis · p. 4

Data analysis

• Ninety-two percent of users have at most ten total actions, while the most active users interact with roughly a hundred papers and successful papers attract hundreds of users.Several prominent authors each contributed more than 15 papers in the sample.Source: Data description and analysis · p. 4

Interaction weighting

• Upload, download, and abstract-view interactions receive weights Wup = 1, Wdown = 0.1, and Wview = 0.05, respectively.These weights distinguish the relative importance assigned to the three actions in the ranking input.Source: Results on real data · p. 4

Paper rankings

• Unlike popularity-driven rankings, QRC avoids letting the highly active forum editor dominate paper scores, although its selected papers are generally younger.Popularity and biHITS favor older or highly downloaded papers, while QRC’s normalization prevents the editor’s weight from dominating.Source: Results on real data · p. 5
• QRC outperforms the other evaluated methods at λ = 0.57, with significance at 0.02 except EigenRumor at 0.08.Table 2 summarizes mean and standard error for submission day, downloads, citation count, and journal impact factor across top-20 paper rankings.Source: Results on real data · p. 5, Table 2 · p. 5
• At λ = 0.57, QRC’s top 20 papers reach an average impact factor of 3.8 and citation count of 34.Both metrics peak near λ = 0.57, while the average number of downloads eventually falls below 25% of the QR1 value as λ increases.Source: Results on real data · p. 5, Figure 3 · p. 5
• QRC mitigates time bias: increasing λ raises the average submission day of its top 20 papers from 375 to 519.The paper also reports that QRC’s selected papers are generally younger than those chosen by other methods, so their citation counts may underestimate importance.Source: Results on real data · p. 5

Author rankings

• QRC’s top 10 authors have mean h-index 41 ± 11, compared with 4 ± 2 for EigenRumor’s top 12 authors.The QRC ranking includes reputed econophysics researchers, whereas EigenRumor’s compared authors all have identical credit.Source: Results on real data · p. 6, Table 3 · p. 6

Discussion. –

• QRC’s top papers attract significantly more citations and come from venues with higher average impact factors than papers selected by other algorithms.They are also younger than papers selected by bipartite HITS and have considerably fewer downloads than papers selected by the other algorithms.Source: Discussion. – · p. 6
• QRC combines quality of papers, reputation of users, and credit of authors in a single reputation algorithm for scientific online communities.Source: Discussion. – · p. 6
• QRC’s top authors have on average substantially higher h-index than top authors found with other algorithms.Source: Discussion. – · p. 6
• QRC can apply beyond scientific communities when shared quality perceptions emerge, quality induces popularity, and items have multiple authors.Divided communities may violate the shared-perception condition, while imperfect quality–popularity causality can limit rankings in some domains.Source: Discussion. – · p. 6
• QRC’s results are community-specific, while open issues include parameter choices, survey-based evaluation, metric gaming, time decay, and rankings over longer data spans.The authors also note that QRC favors active authors and those collaborating with credible authors, motivating further study of gaming and robustness.Source: Discussion. – · p. 6