Fine-Mapping the Source of a Citation Burst Across Research Communities

When a scientific idea suddenly takes off, burst detectors can say that a concept surged, but not which earlier paper set it off. The cause is hidden because the papers behind a new idea are almost always cited together. A paper that adopts a new technique typically cites the paper that introduced it alongside three or four companions: the benchmark it was tested on, a popular follow-up, a survey. Every one of those papers bursts at the same moment, and none can be told apart from citation counts alone. Statistical geneticists face the same problem when locating disease-causing mutations. Neighboring DNA variants are inherited together in blocks, so a whole block lights up in an association scan. The field's fix, cross-ancestry fine-mapping, is to compare populations whose blocks are broken up in different places. A true causal variant stays associated with the trait in every population, while its passengers drop out wherever the block has been broken.

We propose to carry that fix over to scientometrics. Different research communities, such as NLP, computer vision, and biomedical informatics, cite the same cluster of papers with different co-citation patterns. A follow-up paper may be cited in tandem with the original in one field and cited alone in another. We treat each community as a "population" and the co-citation correlation among candidate papers as the analogue of genetic linkage. We then fit a cross-community fine-mapping model whose output is a small, calibrated credible set: a group of papers that contains the true source of a concept burst with stated probability, for example 95%. The core claim is testable. If co-citation patterns differ enough across communities, adding a second community should shrink the credible set the way adding African-ancestry samples shrinks genetic credible sets, and the shrinkage should scale with how different the co-citation patterns are.

Core Claim

Attributing a burst to its source is a sparse variable-selection problem, and its main obstacle is correlated candidates, not too few data points. The rest of the design follows from that.

The unit of analysis is a citing paper $i$ in community $m$ that is published during a burst window. That window comes from Kleinberg's state model, a hidden-Markov detector that flags stretches where events arrive much faster than the background rate; the research file describes it as the engine behind CiteSpace. The outcome $y_{mi}$ records whether paper $i$ uses the bursting concept. The predictors $x_{mi}$ are indicators for which of $P$ candidate papers it cites. Marginal association statistics per candidate are collected into a vector $z_m$ and modeled exactly as multi-population genetic fine-mappers model summary statistics:

$$z_m \sim \mathcal{N}\big(R_m \Lambda_m \beta_m,\; R_m + c_m I\big),$$

Here $R_m$ is the co-citation correlation matrix among candidates in community $m$ (the analogue of the linkage matrix). $\Lambda_m$ is a diagonal matrix of community sample sizes. $\beta_m$ is the vector of true per-candidate effects. $c_m$ absorbs community-wide inflation, such as recency bias or a citation clique, that pushes up all association statistics at once. The effect vector is written as a sum of $L$ single effects, the device introduced in SuSiE. Each single effect picks exactly one source paper, and those picks are shared across communities, while the effect sizes may vary by community.

The inflation term $c_m$ is estimated the way XMAP estimates confounding. In genetics, the intercept of a regression of squared association statistics on per-variant correlation scores measures stratification. We regress squared statistics for all papers in the community on each paper's total co-citation score, and the intercept estimates $c_m$.

Research Program

The method is a direct reuse. We take SuSiE's single-effect decomposition and XMAP's variational expectation-maximization, a fast approximate posterior fit whose cost grows linearly in the number of causal effects rather than combinatorially. We run them on co-citation matrices built from S2ORC, which resolves 380.5 million citation links across 81.1 million papers. Communities are defined by S2ORC field-of-study labels and, inside NLP, by venue.

We make one modeling extension that genetics itself flags as open. Both XMAP and the red-blood-cell study of Hodonsky et al. note that some causal variants are private to one ancestry. The scientometric analogue is common: a concept can be independently reinvented in a second field. We therefore add a per-community spike-and-slab indicator on each single effect. A spike-and-slab prior is a mixture of "exactly zero" and "freely sized" components, and BurstGraph uses the same prior to separate sparse bursty edges from routine ones. Here it lets a source be shared or community-specific.

Two constraints from prior work shape the preprocessing. Bollmann and Elliott show that falling citation age in the ACL Anthology is largely a denominator effect of publication growth, with a 42% year-on-year increase in papers between 2017 and 2018. Singh et al. show a −0.72 correlation between citation volume and citation-age diversity. Both mean that raw citation rates are confounded by volume and recency. We therefore compute statistics within publication-year strata and include candidate age as a covariate, in addition to $c_m$.

The model can optionally use per-candidate "annotations," as PAINTOR does with functional genomic marks. Our annotation is the fraction of a candidate's incoming citations that sit in Methods sections, which S2ORC's section-linked citation mentions make available.

Baselines

The naive reference is the CiteSpace practice in studies such as the glaucoma biomechanics bibliometric analysis, which flagged optical coherence tomography and finite element analysis as the strongest bursts. That practice names the candidate with the highest Kleinberg burst strength as the driver. The second reference is the genetic "lead variant" analogue: rank candidates by marginal association $|z|$ with concept use in the pooled corpus. Both are run on exactly the same candidate sets and burst windows. Neither produces calibrated uncertainty, so we turn each into a set by taking its top-$k$, with $k$ matched to our credible-set size, and compare coverage at equal set size.

The strongest existing method is single-population SuSiE run on all communities pooled into one corpus. It already handles correlated candidates and yields credible sets; it lacks only the cross-community contrast. The ablation that isolates our key ingredient is the reverse comparison: the full model with communities collapsed versus kept separate, holding the variational algorithm, $L$, priors, and iteration budget fixed. A second ablation drops the inflation term $c_m$. A third drops the community-specific indicators.

There is no reported result found on Lacuna for burst source attribution under any metric, so no published number exists to beat. The nearest reported figures concern related tasks: SuSiE's speedup of up to 4,000 times over CAVIAR, XMAP identifying 149 putative causal variants for LDL cholesterol versus about a third as many for SuSiE, and CPB reaching 92.4% accuracy and 0.934 AUC on next-window burst timing in Weibo cascades. The last addresses when a burst happens, not what caused it, so it is context rather than a competitor.

Beating the field means two things. First, at matched 95% coverage, credible sets must be strictly smaller than pooled SuSiE's. Second, at equal top-$k$, recall of the silver-labeled source must be higher than the marginal and burst-strength baselines.

Why not just…?

Why not just read the citation sentence, which often says "introduced by X"? This is the most serious threat, and we use it as the evaluation signal rather than ignoring it. We extract explicit attribution phrases from S2ORC citation contexts as silver labels and withhold them from the model. Many adopting papers use a concept without such a phrase, or attribute it to a popular follow-up rather than the origin. The decisive comparison restricts evaluation to bursts where the attribution text is split across several candidates. If a text-majority-vote attributor matches our model there too, the statistical machinery adds value only for unlabeled corpora, and we will report that plainly.

Why not just take the earliest-published candidate? Temporal precedence fails when the true source is a later paper that popularized or repaired an earlier idea, which is exactly the burst-triggering case. It also cannot represent concepts with two independent sources. We include an "earliest candidate within the credible-set-size budget" baseline so the experiments measure this directly.

Why not just pool all communities into one bigger corpus? Pooling averages the co-citation matrices. The places where a passenger decouples from the source in one field get diluted by fields where they remain tied, and community-level inflation gets mixed into the signal. In genetics, the analogous contrast is single-population SuSiE versus XMAP, and the cross-ancestry MTFR1 credible set in the fetal brain xQTL atlas shrank from about 50 variants to 3–5. The pooled-versus-separate ablation tests whether a comparable gain exists for citations. If it does not, the transfer fails.

Proposed Experiments

The first experiment is a semi-synthetic benchmark built on real co-citation structure, mirroring how SuSiE was benchmarked on real linkage matrices from GTEx. We take real reference matrices from three or more communities and plant one or two source papers. We simulate concept use through a logistic model, with a tunable community-wide confounder and an option for a community-private source. Metrics are credible-set coverage, set size, calibration of posterior inclusion probabilities, and precision-recall of source recovery as the confounder grows.

The second experiment uses real bursts. Concept bursts are detected by Kleinberg's model over S2ORC titles and abstracts across NLP, vision, and biomedical text. Evaluation uses the held-out text-derived silver labels, plus finding-level links from the finding-citation graph for the biomedical subset. A small, fully inspectable case study reruns the glaucoma corpus to ask which papers drove the optical coherence tomography and finite element analysis bursts.

The decisive check is a dose-response prediction. Across bursts, the reduction in credible-set size gained by adding a community should increase with the divergence between that community's co-citation matrix and the others' (measured, for example, as the Frobenius distance between correlation matrices). It should be near zero when the matrices agree. If that relationship holds and the method beats the text-ambiguous baseline, the genetic analogy is doing real work. If it does not, the result shows that co-citation correlation does not behave like linkage, and we will report it as a negative result.