Who ends up at the center of the hallway conversations at a scientific conference? One might expect the people who are famous online, but the evidence says otherwise. When researchers wore proximity badges at four computer science conferences, the people most central in the face-to-face contact network showed no correlation with online prominence measures such as the Google PageRank of their homepage or their ResearchGate score (Is Web Content a Good Proxy for Real-Life Interaction?). That paper documents this online-offline gap but does not explain it. We propose an explanation that can be tested. Offline prominence is not a property of the person. It is a property of the person's topical position relative to the particular crowd in the room. More precisely, people become central in face-to-face networks when their research spans several of the topic clusters that are present at that event, which is what we mean by "brokers." Being typical of the event, or being well known on the web, should matter less.
The test uses three tools. Topic modeling groups text into themes. Here it is run on the pooled homepage and publication text of one event's attendees, which yields that event's topic clusters. Graph centrality ranking scores nodes by their position in a network. Here it is applied to a graph linking people to the event's topics, and it produces a brokerage score for each attendee. Finally, the online-offline gap supplies the ground truth: badge-recorded face-to-face centrality. The decisive test is within-person. Some researchers attended several of the four conferences. If brokerage is the mechanism, a person's face-to-face centrality should rise at events where their topics bridge that event's clusters and fall where they do not, even though their online prominence stays fixed.
Core Claim
Among conference attendees, event-relative topical brokerage predicts face-to-face eigenvector centrality better than three alternatives: online prominence, coauthorship degree, and topical typicality (closeness to the event's average interest). It also tracks within-person changes in centrality across events. The novelty is not a better text-similarity function. It is the claim that the online-offline centrality disconnect arises because online ranking metrics are computed on a global, contagion-shaped graph, while physical prominence is computed on a small, event-specific population whose topical composition changes from event to event.
This builds on three observations in the file. First, the online-offline gap literature shows that online prominence and web-derived networks poorly track casual encounters. The Jaccard overlap between face-to-face and ResearchGate networks is only 0.15 to 0.32 (Kibanov et al.). Separately, observed online topology can itself be an artifact of spreading processes rather than a picture of underlying ties (Potential Networks, Contagious Communities). Both motivate looking away from online topology. Second, the same Kibanov study found that informal homepage text predicts face-to-face contact among non-coauthors better than publication abstracts, so text carries signal that topology misses. That study used bag-of-words cosine similarity between pairs and never asked about centrality. Third, degeneracy-based graph methods separate a text's cohesive core from generic hub terms that span unrelated threads (GoWvis). That paper motivates our choice of salient terms, but it works on single documents and never on groups of people.
Research Program
For each event in the CONFERATOR data (LWA 2010, 2011, 2012 and ACM Hypertext 2011), we collect every attendee's homepage text and DBLP-linked abstracts. Each person's text is represented by its salient core rather than its full bag of words. We build a graph-of-words for that person, meaning a network in which words are nodes and edges join words that appear near each other. We then keep the terms in the main k-core, the largest subgraph in which every word is linked to at least k others (GoWvis). This removes generic hub terms that would otherwise make everyone look like a broker. The retained text is embedded with a sentence encoder.
The event's topics are found by clustering the pooled embeddings of that event's attendees with a BERTopic-style pipeline. This step has a known failure mode. On short text, density-based clustering discarded about 74% of documents as outliers (Experiments on Generalizability of BERTopic), and brokers are exactly the people most likely to fall between clusters. We therefore use soft membership weights $w_{ic}$, the similarity of person $i$ to topic cluster $c$, rather than hard assignment. We then define brokerage as a participation coefficient on the person-topic graph:
$$B_i = 1 - \sum_{c} \left( \frac{w_{ic}}{\sum_{c'} w_{ic'}} \right)^2,$$
which is near zero when a person's interest sits in one of the event's clusters and grows as it spreads evenly across them.
We also compute betweenness of person $i$ on a person-person graph whose edges are weighted by shared topic mass. This gives a second, path-based brokerage measure.
Typicality $T_i$ is the cosine similarity of person $i$'s embedding to the event centroid. It is the natural rival: it says central people are simply the most on-topic attendees.
Every measure is compared against badge-derived face-to-face eigenvector centrality at the same event.
Baselines
The naive reference point is online prominence: homepage Google PageRank and ResearchGate score used directly as predictors of face-to-face centrality. Kibanov et al. report a "surprising lack of correlation" between these and face-to-face eigenvector centrality, but the file gives no correlation value. There is no reported result on Lacuna for a numeric online-prominence correlation, so we recompute it on identical attendee sets. Brokerage beats this baseline if its Spearman correlation with face-to-face centrality is significantly higher, judged by bootstrap confidence intervals.
The strongest existing method is the Kibanov bag-of-words TF-IDF cosine approach, which predicts pairwise face-to-face contact. We aggregate its pairwise scores into a per-person expected degree and also compute eigenvector centrality on the TF-IDF similarity graph. The file reports the network overlap (Jaccard 0.15 to 0.32) but no link-prediction score, so there is no reported result on Lacuna for TF-IDF face-to-face prediction accuracy. We rerun it on the same text and the same attendees, with the same budget for sweeping vocabulary thresholds.
A structural baseline uses DBLP coauthor degree within the event. Kibanov et al. found that people with many coauthors (more than 67) and homepage PageRank above 3 are the "good proxies."
The most dangerous competitor is typicality $T_i$, built from the same embeddings. It is dangerous because it shares every text-processing component with our method and differs only in the hypothesis.
Two ablations isolate the key ingredients. The first removes event-conditioning by computing topics once over all four events. If brokerage survives this change, the event-relative claim fails. The second replaces the k-core salient terms with full text, to measure how much hub-term removal matters.
All methods are tuned with the same small grid, and all use nested leave-one-event-out validation.
Why not just…?
Why not just use TF-IDF or LDA topics? These run as ablations of the text representation, not as a separate hypothesis. Evidence that bag-of-words methods fall short on sparse informal text is indirect but consistent. LDA coherence was 0.031 versus 0.091 for BERTopic on short text (de Groot et al.), and TextRank beat TF-IDF on top-1 interest tags, 67.3% to 59.6% (Wu et al.). If TF-IDF-based brokerage performs as well as embedding-based brokerage, the brokerage claim still stands and becomes cheaper to apply.
Why not just use network topology, such as coauthorship degree? Coauthorship accurately proxies strong ties (86% on average become ResearchGate links) but not casual encounters (Kibanov et al.). We test whether brokerage adds predictive power after controlling for coauthor degree, using partial correlation and Quadratic Assignment Procedure (QAP) permutation tests for the pairwise version.
Why not just measure how on-topic a person is for the event? This is a real threat. On-topic attendees may simply attend more sessions. The two hypotheses make different predictions for people who are highly typical but concentrated in one cluster, versus people who are less typical but spread across clusters. We test brokerage's added explained variance over typicality and report honestly if it is zero.
Why not just train a supervised classifier? With 66 fully observed researchers there is too little labeled data. A fine-tuned model like the Longformer used for Reddit in-group detection (Minot et al.) would also give no mechanism to test.
Proposed Experiments
The first experiment is cross-sectional. At each event, we compute Spearman correlations between each predictor and face-to-face eigenvector centrality, and we fit a regression containing brokerage, typicality, coauthor degree, and online prominence together.
The decisive experiment is within-person. For researchers present at two or more events, we correlate each person's change in brokerage with their change in face-to-face centrality rank. Online prominence is nearly constant across these conference years and cannot explain such changes. Typicality can, which is why it remains the key competitor. We use person-level permutation to account for the small sample.
A pairwise check asks whether two people bridging the same pair of clusters meet more often than predicted by TF-IDF similarity. This is tested on non-coauthor pairs, the regime where Kibanov et al. found text most informative.
Statistical power is the main risk. With a few dozen repeat attendees, only moderate effects will be detectable, and we will state the minimum detectable correlation before running the analysis.
For replication, the SocioPatterns four-conference data offers badge records and surveys at WS16, ICCSS17, ECSS18, and ECIR19 (Génois et al.). A replication there depends on linking attendees to public publication text under the dataset's consent terms. Its Big Five survey measures also let us check that brokerage is not simply a stand-in for extraversion.
The outcome is falsifiable in both directions. Brokerage that beats typicality and tracks within-person change would explain why online and offline prominence diverge. A null result would show that topical position, however it is measured, does not account for the gap.
References
- Is Web Content a Good Proxy for Real-Life Interaction?: A Case Study Considering Online and Offline Interactions of Computer Scientists
- Potential Networks, Contagious Communities, and Understanding Social Network Structure
- GoWvis: A Web Application for Graph-of-Words-based Text Visualization and Summarization
- Experiments on Generalizability of BERTopic on Multi-Domain Short Text
- Automatic Generation of Personalized Annotation Tags for Twitter Users
- Distinguishing In-Groups and Onlookers by Language Use
- Combining Sensors and Surveys to Study Social Interactions: A Case of Four Science Conferences