Citation Pattern Noise in Research

• Citation pattern noise is the intra-author inconsistency in citation decisions relative to an ideal knowledge flow standard.
• It decomposes into stable noise reflecting systematic habits and occasion noise representing transient, random errors.
• The concept challenges count-based metrics and motivates enhanced citation decision hygiene to improve research assessment.

Searching arXiv for recent and foundational papers on citation pattern noise, citation fidelity, and citation structure. Citation pattern noise is a concept in citation analysis that denotes the inconsistency of citation decisions within the work of a single author across different papers and citation occasions. In the framework proposed in “Citation accuracy, citation noise, and citation bias: A foundation of citation analysis” (Bornmann et al., 18 Aug 2025), it is one component of a broader theory of citation noise, alongside citation level noise, and is defined relative to a normative ideal in which a citation is accurate only when a concrete knowledge flow from the cited paper to the citing paper has taken place. The concept therefore treats citation records not merely as counts, but as noisy observations of underlying intellectual dependence. Related literatures on citation fidelity and citation-network structure illuminate complementary aspects of the same problem: whether citations faithfully transmit scientific claims, and whether reference lists are coherent, idiosyncratic, or brokerage-like in their induced network topology (Chen et al., 27 Feb 2025, Shi et al., 2010).

1. Conceptual definition and scope

In the cited framework, citation noise is the undesirable variance in citation decisions, and citation pattern noise is its within-author component. More precisely, it concerns how one and the same author’s citation behavior varies across different papers and across different candidate cited papers, even when citation is supposed to reflect knowledge flow (Bornmann et al., 18 Aug 2025).

This definition distinguishes citation pattern noise from a more generic complaint that citations are “messy.” The concept is not about aggregate citation counts fluctuating across fields or years, nor primarily about whether some authors are generally more careful than others. Instead, it concerns intra-author inconsistency: an author may be relatively accurate on average and yet still exhibit substantial instability in which papers are cited, omitted, or misattributed across different contexts.

The normative baseline is citation accuracy. In this scheme, a citation is accurate if another paper is cited exactly when a knowledge flow has taken place. The framework explicitly allows four logical possibilities: correct positive, correct negative, incorrect positive, and incorrect negative. Both over-citation and under-citation are therefore treated as errors relative to the knowledge-flow norm (Bornmann et al., 18 Aug 2025).

A central implication is that citation pattern noise is not reducible to raw citation frequency. Two papers may receive the same number of citations, yet one may be cited through relatively consistent knowledge-flow judgments while the other is cited through inconsistent, context-dependent, or accidental decisions. This suggests that count-based indicators can obscure substantial variation in the reliability of citation acts.

2. Formalization within the citation-noise framework

The formal definition of author-specific citation pattern noise in the framework is

$$OPN_i = \frac{1}{n_i}\sum_{j=1}^{n_i}(PE_i - PE_{ij})^2$$

where $n_i$ is the number of citing papers authored by author $i$, $PE_i$ is the author-specific average error rate of author $i$, and $PE_{ij}$ is the error proportion in citing paper $j$ by author $i$ (Bornmann et al., 18 Aug 2025).

Overall citation pattern noise is then defined as the weighted average of these author-specific quantities:

$OPN = \frac{\sum_i n_i \times OPN_i}{\sum_i n_i}.$

This formalization makes clear that citation pattern noise is a variance term. It measures dispersion around an author’s own average error propensity rather than deviation from a system-wide average. For that reason, it is conceptually distinct from citation level noise, which is the between-author variance in erroneous citation decisions and is given by

$$OLN = \sqrt{\frac{\sum_i n_i(PE - PE_i)^2}{\sum_i n_i}}$$

with $n_i$0 denoting the proportion of erroneous citations in the entire citation system (Bornmann et al., 18 Aug 2025).

The distinction is substantive. Citation level noise asks whether some authors systematically make more citation errors than others. Citation pattern noise asks whether the same author responds inconsistently across citation opportunities. The first is a cross-sectional property of authors; the second is an interactional property of authors, papers, and occasions.

The paper illustrates these constructs using a fictitious social citation system with 10 citing papers written by 3 citing authors and 5 cited papers. In that system, the table records realized citations, accurate citations, erroneous citations, author-specific and paper-specific proportions, and the resulting overall level noise and overall pattern noise. The purpose of that construction is to show that a citation system can appear acceptable at the aggregate level while still containing substantial within-author inconsistency (Bornmann et al., 18 Aug 2025).

3. Stable pattern noise, occasion noise, and generating mechanisms

The framework further decomposes citation pattern noise into two components: stable citation pattern noise and citation occasion noise. Stable citation pattern noise is the systematic, repeatable part of within-author variability. It arises when an author has persistent tendencies to cite certain kinds of work in some contexts but not in others. Citation occasion noise is the random part, generated by transient circumstances that affect whether a citation is made in a particular instance (Bornmann et al., 18 Aug 2025).

The examples provided are deliberately concrete. Stable pattern noise includes cases such as an author who tends to cite only a few sources but especially likes papers about statistical methods, or an author who cites many papers but mostly papers by European authors. Occasion noise includes cases such as an author who happens to notice a passage and cites it on that occasion but would probably not have cited it at another time, or an author who usually cites extensively but cites less in one period because of a heavy teaching workload (Bornmann et al., 18 Aug 2025).

This decomposition matters because not all noise is random in the same sense. Stable pattern noise is systematic within authors even if it is undesirable relative to the knowledge-flow norm. Citation occasion noise is explicitly characterized as random error. A plausible implication is that interventions aimed at reducing citation noise must distinguish durable habits from transient perturbations; training and norms may address the former, while aggregation may partially absorb the latter.

The paper identifies several mechanisms that can generate citation pattern noise: interaction between different paper characteristics and author tendencies, superficial or incomplete reading, temporary conditions, citation habits and style, external influence such as reviewer or editor pressure, coercive citation, access limitations, and database or reference errors (Bornmann et al., 18 Aug 2025). These mechanisms jointly imply that pattern noise is not a single pathology but a family of departures from consistent knowledge-flow attribution.

4. Relation to citation fidelity and the transmission of scientific claims

A closely related empirical literature operationalizes citation noise through citation fidelity rather than through explicit error-variance decomposition. “The Noisy Path from Source to Citation: Measuring How Scholars Engage with Past Research” (Chen et al., 27 Feb 2025) defines citation fidelity as the degree to which a citing sentence preserves the scientific information of the original claim in the cited paper. In that study, low fidelity is the noisy part of the citation channel.

The operational measure is a sentence-level scientific information change score on a 1–5 scale: 1 indicates completely different scientific claims, 2 mostly different, 3 somewhat similar, 4 mostly the same, and 5 completely the same. Using full texts, the study builds a pipeline over the Semantic Scholar Open Research Corpus, identifies reporting citation sentences, extracts candidate result and conclusion sentences from cited papers, and assigns each citation pair the highest-scoring match between the citing sentence and any candidate sentence in the cited paper. The analysis covers approximately 13 million citation sentence pairs, and a random sample of about 10k fidelity scores is reported as approximately normally distributed around 3.5 with a full range from 1.0 to 5.0 (Chen et al., 27 Feb 2025).

The study’s predictive findings are also relevant. Citation fidelity is higher when cited papers are more recent and intellectually close, more accessible, and when the first author has a lower H-index and the author team is medium-sized. Additional appendix findings report higher fidelity in Biology and Medicine than in Physics and Computer Science, higher fidelity for review articles than for journal articles and conference papers, lower fidelity as the cited paper’s citation count increases, higher fidelity for longer citation contexts until the effect levels off beyond about 100 words, and lower fidelity when the same source is cited repeatedly within a paper (Chen et al., 27 Feb 2025).

Its quasi-experimental “telephone effect” is particularly pertinent. In matched comparisons, papers that cite an original source together with an intermediary paper that itself cites the original have fidelity to the original that is 0.06 lower than papers citing the original directly, and the intermediary’s own fidelity to the original is positively correlated with the later paper’s fidelity (Chen et al., 27 Feb 2025). This does not formally measure citation pattern noise as defined in (Bornmann et al., 18 Aug 2025), but it demonstrates a concrete transmission mechanism by which noisy citation behavior can propagate across citation chains.

5. Citation-pattern structure in network analysis

A separate but adjacent literature studies citation patterns not as within-author error variance but as the network structure of a paper’s reference neighborhood. “Citing for High Impact” (Shi et al., 2010) introduces citation projection graphs, where for a focal paper $n_i$1, $n_i$2 is the subgraph induced by the papers cited by $n_i$3, and $n_i$4 is the graph induced by those cited papers plus $n_i$5 itself. This approach asks how the cited papers cite one another.

Within that framework, three archetypes are described. The idiosyncratic citer draws on references that are sparse and weakly connected. The within-community citer draws on a dense, clustered, cohesive reference set. The brokerage citer draws on multiple communities linked by bridging papers. Six metrics characterize these structures: graph density $n_i$6, clustering coefficient $n_i$7, connectivity as the fraction of nodes in the largest weakly connected component $n_i$8, maximum betweenness $n_i$9, betweenness of the focal paper $i$0, and network constraint of the focal paper $i$1 (Shi et al., 2010).

This network-structural usage does not coincide with the formal definition of citation pattern noise in (Bornmann et al., 18 Aug 2025). Nonetheless, the two perspectives intersect. High $i$2, low connectivity, and low clustering indicate that the focal paper is the only bridge among otherwise disconnected references, a pattern the study associates with idiosyncratic or random citation neighborhoods. In natural and social science, low-impact papers tend to exhibit such scattered structures, whereas medium-impact papers more often show narrow, coherent, discipline-focused citation patterns, and high-impact papers more often exhibit structured crossing-community brokerage. In computer science, by contrast, high-impact papers tend to be more focused, with higher density, clustering, connectivity, and constraint and lower betweenness of $i$3 (Shi et al., 2010).

This suggests that the phrase “citation pattern” has at least two analytically distinct meanings in the literature. In one usage, it denotes inconsistency in citation judgment relative to knowledge flow. In another, it denotes the topology of a paper’s local citation environment. The two are not equivalent, but both are concerned with whether citation behavior is structured, coherent, and interpretable rather than scattered or noisy.

6. Consequences for bibliometrics, evaluation, and reform

The principal significance of citation pattern noise is methodological. Citation-based indicators are widely used as proxies for knowledge flow, usefulness, quality, and impact. If citation acts contain substantial within-author inconsistency, then citation rates, times cited, paper rankings, and more specialized indicators can become unstable measures of the constructs they are intended to capture (Bornmann et al., 18 Aug 2025).

The framework argues that noise is especially damaging at low levels of aggregation. Large aggregates can partially average out random error under a wisdom-of-crowds logic, but individual papers, authors, and small groups are much more vulnerable. The paper therefore recommends aggregation as one partial strategy, while stressing that aggregation reduces noise more than bias and does not repair inaccurate decisions at the individual level (Bornmann et al., 18 Aug 2025).

The broader program it proposes is “citation decision hygiene.” Recommended measures include guidelines for accurate citations, training researchers to cite only when knowledge flow has occurred, reviewer and editor involvement in detecting citation errors, citation justification tables documenting why each work is cited, correction mechanisms for published citation mistakes, and cautious use of AI as a pre-publication aid for flagging potentially missing or misused citations. The paper emphasizes that AI cannot autonomously judge knowledge flow, though it may support human checking (Bornmann et al., 18 Aug 2025).

The relevance for research assessment reform is explicit. The paper recommends that the Coalition for Advancing Research Assessment (CoARA) and related initiatives treat citation noise as a foundational issue. The argument is not that citation-based assessment must be abandoned in every form, but that the reliability and validity of citation analysis depend on improving the quality of citation decisions themselves (Bornmann et al., 18 Aug 2025).

Taken together, the available literature presents citation pattern noise as a problem of signal degradation in scholarly communication. In the strict formal sense, it is the within-author variance of citation errors. In related empirical work, it appears as loss of fidelity in the transmission of claims and as idiosyncratic structure in reference neighborhoods. Across these formulations, the common concern is that citations are not homogeneous tokens of intellectual influence. They are judgmental, context-sensitive, and sometimes distorted traces of knowledge flow, and the analysis of science must account for that variability if citation data are to remain credible evidence.