arxiv: 2604.23699 · v1 · submitted 2026-04-26 · 💻 cs.DL · cs.LG

Recognition: unknown

Beyond coauthorship: semantic structure and phantom collaborators in transportation research, 1967--2025

SeongJin Choi

Authors on Pith no claims yet

Pith reviewed 2026-05-08 04:58 UTC · model grok-4.3

classification 💻 cs.DL cs.LG

keywords coauthorship networkssemantic similaritycollaboration predictiontransportation researchphantom collaboratorsresearch communitiestopic communitiesnetwork analysis

0 comments

The pith

Authors close in semantic space but far in the coauthorship graph become real collaborators at rates 16 to 33 times above baseline.

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

The paper builds a semantic and structural map of transportation research from over 120,000 papers across decades. It shows that communities formed by paper content overlap only weakly with communities formed by who has actually written papers together. The central advance defines phantom collaborators as pairs who rank as top semantic neighbors yet sit at least three hops apart in the coauthorship network. A hold-out test demonstrates these pairs form new coauthorships in later years at dramatically elevated rates that increase sharply with greater semantic similarity.

Core claim

Phantom collaborators are pairs of authors who are top-K semantic neighbors based on combined paper embeddings, TF-IDF, and venue projections, yet remain at least three hops apart in the coauthor graph. When trained on data through 2019, these pairs coauthor new work in 2020-2025 at 16 to 33 times the rate of random, popularity-weighted, and same-venue baselines, with conversion rates rising monotonically by a factor of 68 from the lowest- to highest-similarity buckets.

What carries the argument

Phantom collaborator: an author pair that is nearest-neighbor in semantic embedding space of their papers but separated by three or more steps in the coauthorship graph.

Load-bearing premise

The embeddings and projections measure collaboration-relevant similarity without being driven by unmeasured factors such as shared geography, career stage, or external events.

What would settle it

Repeating the hold-out test on publications after 2025 and finding that phantom pairs no longer show elevated coauthorship rates compared with the same baselines.

Figures

Figures reproduced from arXiv: 2604.23699 by SeongJin Choi.

**Figure 1.** Figure 1: Corpus growth, author-team size, and Lotka-style productivity. consistent with the broader science finding that team size drifts upward as data sharing, co-development, and crossinstitutional collaboration become cheaper (Newman, 2001). The IEEE venues (T-ITS and T-IV) accelerate sharply after 2018, reaching a mean above five authors per paper and by 2024 approaching six, as ML-heavy submissions with a fi… view at source ↗

**Figure 2.** Figure 2: Two views of the coauthor-network largest connected component (28,005 authors, top-12 Leiden communities colored). 1980 1990 2000 2010 2020 Cutoff year (edges with first collaboration ≤ y) 0 20 40 60 80 100 Largest-component fraction (%) Sun & Rahwan (2017): 58% view at source ↗

**Figure 3.** Figure 3: Giant-component fraction versus cutoff year. choice modeling, pavement and materials engineering, and, more recently, connected and automated vehicles. These groupings broadly match the editorial scopes of TR Parts E, B, A, D, and C, respectively. 5.3. Bridge edges between communities view at source ↗

**Figure 4.** Figure 4: Top-100 bridge edges, aggregated by community pair. Leiden communities has grown from ∼ 20 to 172, and the Lotka exponent has flattened from 𝛼 ≈ 2.6 to 𝛼 ≈ 2.28 — all consistent with the Newman (2001) characterization of a mature, tightly connected scientific community. S. Choi: Preprint submitted to Elsevier Page 9 of 38 view at source ↗

**Figure 5.** Figure 5: Author topic space (UMAP projection of the hybrid embedding), colored by semantic-Leiden community. 6.4. Why the production embedding is the hybrid We ship the three-component hybrid (whitened SPECTER2 plus concept TF-IDF plus venue LDA) rather than raw SPECTER2, even though a small ablation we ran on this corpus gives raw a roughly one- to two-percentagepoint edge on phantom-collaborator precision at 𝐾 =… view at source ↗

**Figure 6.** Figure 6: Pairwise cosine similarity, before and after Arora et al. (2017) whitening view at source ↗

**Figure 7.** Figure 7: Row-normalized co-occurrence between semantic communities (rows) and top-22 coauthor communities (columns). are already central. An open, semantic atlas is the natural place to test whether paper-embedding proximity carries collaboration-relevant information that coauthorship structure alone does not encode. If it does, the finding is both a validation of the hybrid embedding built in Sec. 6 and the basis … view at source ↗

**Figure 8.** Figure 8: Precision at 𝐾 for the phantom predictor and three null baselines view at source ↗

**Figure 9.** Figure 9: Relationship between pairwise cosine similarity and the realized future-coauthor rate at 𝐾 = 20. at 𝐾 = 20), confirming that the predictor is not merely re-discovering venue-mediated clustering. The headline finding is therefore robust to the social-distance cutoff: when triadic-closure cases are admitted, raw precision approximately doubles but the multiplicative advantage of semantic similarity over null… view at source ↗

**Figure 10.** Figure 10: Total path length versus net displacement (UMAP units), restricted to the 3,537 authors with ≥ 3 five-year bins. S. Choi: Preprint submitted to Elsevier Page 20 of 38 view at source ↗

**Figure 11.** Figure 11: Exemplar trajectories. 10. The Transport Atlas tool The analyses in Secs. 4–8 are backed by a public interactive atlas at https://choi-seongjin.github.io/ transport-atlas/. The atlas is a static site with six views. Explorer offers a Tabulator.js table of all papers with free-text search, per-venue filtering, and citation-count sorting. Papers by Year overlays the stacked annual counts used in Fig. 1a wit… view at source ↗

**Figure 12.** Figure 12: reports the degree and strength distributions on a log–log scale. Both are heavy-tailed with OLS slope close to −2.3, consistent with the 𝛼 ≈ 2.2 that Newman (2001) found for physics and biomedicine view at source ↗

**Figure 13.** Figure 13: Per-author citation distribution (log–log). F.2. Shortest paths and centrality correlations view at source ↗

**Figure 14.** Figure 14: Shortest-path length distribution inside the LCC. d s c bc bcw pr prw d s c bc bcw pr prw 1.00 0.90 0.27 0.51 0.46 0.55 0.50 0.90 1.00 0.31 0.49 0.44 0.49 0.49 0.27 0.31 1.00 0.39 0.37 0.21 0.23 0.51 0.49 0.39 1.00 0.86 0.44 0.42 0.46 0.44 0.37 0.86 1.00 0.44 0.42 0.55 0.49 0.21 0.44 0.44 1.00 0.87 0.50 0.49 0.23 0.42 0.42 0.87 1.00 −1.00 −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 K endall τ view at source ↗

**Figure 15.** Figure 15: Kendall-𝜏 correlations among author centrality metrics. F.3. Top-30 authors by combined centrality view at source ↗

read the original abstract

We present a semantic-structural atlas of transportation research built from 120{,}323 papers across 34 peer-reviewed journals published between 1967 and 2025, roughly an order of magnitude larger than and a decade beyond Sun and Rahwan's~(2017) coauthorship study. We use OpenAlex and Crossref as open, CC0-licensed data sources, resolve author identity through OpenAlex author IDs, ORCID records, and manual alias resolution, and embed every paper with SPECTER2 with Arora-style whitening concatenated with concept TF--IDF and venue linear-discriminant projections. On this substrate we report three findings. First, Leiden on the author-level semantic k-nearest-neighbor graph yields 23 topic communities that agree only weakly with the 172 coauthor communities (normalized mutual information $0.23$), opening room for a predictive layer that neither source encodes alone. Second, a multiplex Leiden partition combining both edge types recovers 181 communities and localizes where collaboration and topic structure decouple. Third -- the paper's core methodological contribution -- we define \emph{phantom collaborators}, pairs of authors who are top-$K$ semantic neighbors yet $\geq 3$ hops apart in the coauthor graph, and show via a temporal hold-out (training cutoff 2019) that phantom pairs become real coauthors in 2020--2025 at a rate $16$ to $33$ times above random, popularity-weighted, and same-venue baselines, with a $68$-fold monotone gradient between the highest- and lowest-similarity buckets. All artifacts are released as a live, reproducible web atlas at https://choi-seongjin.github.io/transport-atlas/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper defines phantom collaborators via semantic embeddings and shows they predict future coauthorship 16-33x above listed baselines in a large temporal hold-out, but institutional co-location is a plausible unmeasured driver.

read the letter

The core contribution here is the phantom-collaborator idea: pairs that sit close in SPECTER2-plus-TF-IDF space but at least three hops apart in the pre-2019 coauthor graph, then turn out to coauthor at much higher rates in 2020-2025. They run this on 120k transportation papers, use a clean 2019 cutoff for the hold-out, and release the full atlas with code and data. That temporal validation and the scale are the parts that actually move the needle beyond the 2017 Sun and Rahwan baseline they cite. The community-detection sections also show that semantic and structural partitions only weakly overlap, which justifies looking for a predictive layer that uses both signals. The open-data pipeline and the live web atlas are practical pluses that make the work usable by others. The soft spot is exactly the one the stress-test flags. The baselines are random, popularity-weighted, and same-venue, but none condition on institutional or geographic proximity. OpenAlex supplies affiliation strings, and co-location is a known strong predictor of collaboration; semantic embeddings are likely to correlate with it because authors at the same place tend to work on overlapping topics. Without that control, the reported 16-33x lift and the 68-fold similarity gradient could be partly or largely non-semantic. Sensitivity to K and the hop threshold is also not shown in detail, which matters for reproducibility. This is a methods paper aimed at scientometrics and network-science people who want to blend text and graph signals in one applied domain. It is not going to reshape general collaboration theory, but the empirical setup is honest enough and the data release is real. I would send it to peer review so the authors can add a co-location baseline and a few robustness checks; the temporal design already gives it enough grounding to be worth referee time.

Referee Report

2 major / 2 minor

Summary. The paper builds a large-scale atlas of transportation research (120k papers, 1967-2025) using SPECTER2 embeddings concatenated with TF-IDF and venue projections. It reports weak alignment (NMI 0.23) between 23 semantic topic communities and 172 coauthorship communities from Leiden clustering, a multiplex partition, and introduces 'phantom collaborators' (top-K semantic neighbors at least 3 hops apart in the pre-2019 coauthor graph). Via temporal hold-out, these pairs show 16-33x higher rates of becoming real coauthors in 2020-2025 than random, popularity-weighted, and same-venue baselines, with a 68-fold similarity gradient; all artifacts are released as a live atlas.

Significance. If the semantic signal can be isolated, the work provides a scalable, predictive approach to latent collaboration potential that neither coauthorship nor topic structure captures alone, with clear strengths in dataset scale, open CC0 sources (OpenAlex/Crossref), author disambiguation, temporal validation that avoids forward leakage, and the public reproducible atlas. These elements support falsifiable claims and reproducibility.

major comments (2)

[phantom collaborators analysis (third finding)] The temporal hold-out evaluation of phantom collaborators (third finding) compares against random, popularity-weighted, and same-venue baselines but does not condition on institutional or geographic co-location. Since co-location is a documented strong predictor of future collaboration and is plausibly correlated with SPECTER2 embeddings, the reported 16-33x lift and 68-fold gradient may be partly attributable to non-semantic factors rather than the claimed semantic signal alone.
[semantic graph construction and Leiden partitions] The author-level semantic k-nearest-neighbor graph construction (used for both community detection and phantom definition) leaves free parameters (top-K, hop cutoff) and the exact whitening procedure unspecified, with no sensitivity analysis; this affects robustness of the NMI=0.23 result and the phantom rates.

minor comments (2)

The multiplex Leiden partition recovering 181 communities is presented without details on edge weighting or resolution parameter choices.
Clarify how venue linear-discriminant projections are concatenated with SPECTER2 and TF-IDF, and whether they are normalized before kNN construction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the robustness of our findings. We address each major comment below and outline the revisions we will incorporate.

read point-by-point responses

Referee: [phantom collaborators analysis (third finding)] The temporal hold-out evaluation of phantom collaborators (third finding) compares against random, popularity-weighted, and same-venue baselines but does not condition on institutional or geographic co-location. Since co-location is a documented strong predictor of future collaboration and is plausibly correlated with SPECTER2 embeddings, the reported 16-33x lift and 68-fold gradient may be partly attributable to non-semantic factors rather than the claimed semantic signal alone.

Authors: We agree that institutional and geographic co-location represents a plausible confound not fully isolated by the same-venue baseline. Although venue overlap provides a partial proxy (as many transportation journals are regionally concentrated), direct controls would strengthen the isolation of the semantic signal. In the revised manuscript we will add an institutional co-location baseline using OpenAlex affiliation strings (matched at the organization level for authors active pre-2019) and report the phantom collaboration rates after conditioning on this factor. We will also discuss the residual lift as evidence that semantic similarity contributes beyond co-location. revision: partial
Referee: [semantic graph construction and Leiden partitions] The author-level semantic k-nearest-neighbor graph construction (used for both community detection and phantom definition) leaves free parameters (top-K, hop cutoff) and the exact whitening procedure unspecified, with no sensitivity analysis; this affects robustness of the NMI=0.23 result and the phantom rates.

Authors: We acknowledge that the manuscript should have explicitly stated the parameter choices and included sensitivity checks. The reported results use K=10 nearest neighbors, a minimum hop distance of 3, and Arora-style whitening applied to the concatenated SPECTER2+TF-IDF+venue vectors. In the revision we will add a dedicated methods subsection and appendix that (i) fully specifies the whitening implementation and (ii) presents sensitivity tables varying K from 5 to 20 and hop cutoffs from 2 to 4, demonstrating that both the NMI value and the phantom collaboration multipliers remain qualitatively stable (within 10-15% relative change). revision: yes

Circularity Check

0 steps flagged

No circularity: temporal hold-out makes phantom transition rates independent of input definitions

full rationale

The paper defines phantom collaborators from pre-2019 semantic kNN (SPECTER2 + TF-IDF + venue projections) and coauthor graph distance, then evaluates their 2020-2025 coauthorship rates against independent baselines on held-out future data. This out-of-sample test means the 16-33x multipliers and 68-fold gradient are empirical observations, not tautological by construction. No self-citations are load-bearing, no parameters are fitted then renamed as predictions, and the embeddings are externally pre-trained. The derivation chain remains self-contained and falsifiable.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The claim rests on the assumption that semantic embeddings reflect future collaboration potential and on the specific thresholds chosen for neighbors and hops.

free parameters (2)

top-K semantic neighbors
Threshold defining who counts as a semantic neighbor; value not stated in abstract.
hop distance cutoff
Set at >=3 hops in coauthor graph; chosen to separate phantoms from direct collaborators.

axioms (1)

domain assumption SPECTER2 embeddings with Arora whitening plus concept TF-IDF and venue LDA projections capture topic similarity relevant to collaboration
Invoked to define phantom pairs and the similarity gradient.

invented entities (1)

phantom collaborators independent evidence
purpose: Label for semantically close but coauthor-distant author pairs that predict future collaboration
New term introduced to organize the predictive finding; independent evidence supplied by the temporal hold-out rates.

pith-pipeline@v0.9.0 · 5613 in / 1311 out tokens · 45054 ms · 2026-05-08T04:58:23.202943+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 15 canonical work pages

[1]

org/abs/2004.07180

doi:10.18653/v1/2020.acl-main.207. van Eck, N.J., Waltman, L.,

work page doi:10.18653/v1/2020.acl-main.207 2020
[2]

Scientometrics 84, 523–538

Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics 84, 523–538. doi:10.1007/s11192-009-0146-3. Fortunato, S., Hric, D.,

work page doi:10.1007/s11192-009-0146-3
[3]

Physics Reports 659, 1–44

Community detection in networks: A user guide. Physics Reports 659, 1–44. doi:10.1016/j.physrep.2016.09

work page doi:10.1016/j.physrep.2016.09 2016
[4]

Quantitative Science Studies , _month = feb, publisher =

Crossref: the sustainable source of community-owned scholarly metadata. Quantitative Science Studies 1, 414–427. doi:10.1162/qss_a_00022. Jacomy, M., Venturini, T., Heymann, S., Bastian, M.,

work page doi:10.1162/qss_a_00022
[5]

PLOS ONE 9, e98679

ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLOS ONE 9, e98679. doi:10.1371/journal.pone.0098679. Jiang, C., Bhat, C.R., Lam, W.H.,

work page doi:10.1371/journal.pone.0098679
[6]

Transportation Research Part B: Methodological 138, 268–291

A bibliometric overview of transportation research part b: Methodological in the past forty years (1979– 2019). Transportation Research Part B: Methodological 138, 268–291. doi:10.1016/j.trb.2020.05.016. Liben-Nowell, D., Kleinberg, J.,

work page doi:10.1016/j.trb.2020.05.016 1979
[7]

Journal of the American Society for Information Science and Technology , volume =

The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology 58, 1019–1031. doi:10.1002/asi.20591. Lotka, A.J.,

work page doi:10.1002/asi.20591
[8]

UMAP: Uniform manifold approximation and projection for dimension reduction.Journal of Open Source Software, 3(29):861, 2018

doi:10.21105/joss.00861. Modak, N.M., Merigó, J.M., Weber, R., Manzor, F., Ortúzar, J.d.D.,

work page doi:10.21105/joss.00861
[9]

Transportation Research Part A: Policy and Practice 120, 188–223

Fifty years of transportation research journals: A bibliometric overview. Transportation Research Part A: Policy and Practice 120, 188–223. doi:10.1016/j.tra.2018.11.015. Mucha, P.J., Richardson, T., Macon, K., Porter, M.A., Onnela, J.P.,

work page doi:10.1016/j.tra.2018.11.015 2018
[10]

Science 328, 876–878

Community structure in time-dependent, multiscale, and multiplex networks. Science 328, 876–878. doi:10.1126/science.1184819. Newman, M.,

work page doi:10.1126/science.1184819
[11]

Proceedings of the National Academy of Sciences 98, 404–409

The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences 98, 404–409. doi:10.1073/pnas.98.2.404. Priem, J., Piwowar, H., Orr, R.,

work page doi:10.1073/pnas.98.2.404
[12]

Openalex: A fully-open index of scholarly works, authors, venues, institutions, and concepts.arXiv preprint arXiv:2205.01833, 2022

OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv:2205.01833. Singh,A.,D’Arcy,M.,Cohan,A.,Downey,D.,Feldman,S.,2023. SciRepEval:Amulti-formatbenchmarkforscientificdocumentrepresentations, in: Proceedings of EMNLP

work page arXiv 2023
[13]

Sun, L., Rahwan, I.,

doi:10.18653/v1/2023.emnlp-main.338. Sun, L., Rahwan, I.,

work page doi:10.18653/v1/2023.emnlp-main.338 2023
[14]

Transportation Research Part A: Policy and Practice 100, 135–151

Coauthorship network in transportation research. Transportation Research Part A: Policy and Practice 100, 135–151. doi:10.1016/j.tra.2017.04.011. Traag, V., Waltman, L., van Eck, N.,

work page doi:10.1016/j.tra.2017.04.011 2017
[15]

doi:10.1038/s41598-019-41695-z. S. Choi:Preprint submitted to ElsevierPage 24 of 38 Semantic structure and phantom collaborators in transportation research A. Venue ISSNs and coverage Table 9 The 34 venues indexed by the atlas. Coverage is the observed year span in OpenAlex for papers of typejournal-article orconference-paper. Venue Abbr. ISSN (print / on...

work page doi:10.1038/s41598-019-41695-z 1979
[16]

Venues contributing fewer than100papers at fit time are folded into an “other” class, which is why the realized dimension is28rather than33

= 28 components. Venues contributing fewer than100papers at fit time are folded into an “other” class, which is why the realized dimension is28rather than33. Hybrid concatenation.The three𝐿 2-normalized blocks are concatenated with square-root-weighted scaling: whitened SPECTER2 at √ 0.55(768dim), concept TF–IDF at √ 0.30(128dim), and venue LDA at √ 0.15(...

2017
[17]

different

Five-yearbins,minimumtwonon-emptybinsperauthorforclassificationeligibility,restrictedtoauthorswith≥3 bins(≥10-yearpublicationwindow,twotrajectorysegments)fortheheadlinefour-classpartition.Centroidtrajectories aresmoothedonlythroughbinaggregation;notemporalregulariserorkernelsmootherisapplied.Classcutoffs:stayer cutoff𝜏 stay = 15, drifter cutoff𝜏𝜂 = 0.60, ...

2021