Source-linked AI summary
Inheritance patterns in citation networks reveal scientific memes
Tobias Kuhn, Matjaz Perc, Dirk Helbing
TL;DR
Frequent terms need not be interesting scientific memes; the paper characterizes memes by combining occurrence frequency with propagation along the citation graph. Tested on more than 47 million publication records, the score was reported to outperform alternative metrics without linguistic or ontological knowledge or arbitrary thresholds.
Problem
Frequent words may occur often without being interesting in a given scientific field, motivating a measure that distinguishes interesting memes from merely frequent terms.
Source: Results · p. 2Method
The meme score formalizes scientific memes as the product of occurrence frequency and a propagation score measuring alignment with the citation graph.
Source: Results · p. 2, Discussion · p. 6Results
The score was demonstrated on more than 47 million publication records and reported superior to alternative metrics, with accuracy evaluated using annotations and citation-graph randomizations.
Source: Discussion · pp. 6–7Takeaways & Limitations
The score can be calculated exactly and exhaustively on massive databases without arbitrary thresholds or linguistic or ontological knowledge.
Source: Discussion · p. 6The analysis considers only fixed character sequences; adding linguistic and ontological processing may capture memes at more abstract levels.
Source: Discussion · p. 7
Abstract
from arXiv · showhide
Memes are the cultural equivalent of genes that spread across human culture by means of imitation. What makes a meme and what distinguishes it from other forms of information, however, is still poorly understood. Our analysis of memes in the scientific literature reveals that they are governed by a surprisingly simple relationship between frequency of occurrence and the degree to which they propagate along the citation graph. We propose a simple formalization of this pattern and we validate it with data from close to 50 million publication records from the Web of Science, PubMed Central, and the American Physical Society. Evaluations relying on human annotators, citation network randomizations, and comparisons with several alternative approaches confirm that our formula is accurate and effective, without a dependence on linguistic or ontological knowledge and without the application of arbitrary thresholds or filters.
Results
Across APS, PubMed Central, and Web of Science, scientific memes follow similar frequency–propagation distributions; in APS, the meme score outperforms alternative metrics against physics Wikipedia terms.
Results
- Across APS, PubMed Central, and Web of Science, frequency and propagation form a similar downward-sloping band, with top memes near its sparse upper-right edge.The 99.9%-quantile line is notably stable despite underlying values spanning five or more orders of magnitude.Source: Results · p. 3, Figure 2 · p. 4
- Time-preserving randomization reduced nonzero-score terms from ≈1.4 million to 89,356 and erased the universal pattern.Randomized-network median meme scores differed by more than an order of magnitude from original-network scores.Source: Results · p. 3, Figure 2 · p. 4
- Most of the top 50 APS memes are noun phrases, spanning one to three words, despite the score using no linguistic or ontological knowledge.The method also imposes no n-gram length limit or explicit balancing for phrase length.Source: Results · p. 3, Table I · p. 4
- Each annotator classified 86% of meme-score terms as important physics concepts, compared with 25% non-weighted and 19% weighted agreement for random terms.For noun phrases, each annotator classified 86% of meme-score terms that way, with 83% inter-annotator agreement; random-term values were 33% non-weighted and 25% weighted.Source: Results · p. 4, Table II · p. 5
- Meme-score A values were 40.9%–44.8% for δ = 2–10, versus below 22% for every alternative metric; δ = 1 yielded 32.3%.The box plot compares metrics by A, the normalized area under the curve for agreement with physics terms from Wikipedia.Source: Results · p. 5, Figure 3 · p. 5
- Physics Wikipedia titles cluster near the 99.9%-quantile, while chemical formulas have relatively low individual frequency and high propagation scores.The chemical-formula density is highest along the 99.9%-quantile.Source: Results · p. 5, Figure 4 · p. 6
- Top physics memes show bursty time dynamics, and only a few remain at the top of the rankings for extended periods.The authors suggest these bursts might reflect the rise and fall of scientific paradigms and principles of self-organization.Source: Results · pp. 5–6, Figure 5 · p. 6
Discussion
The paper formalizes scientific memes as a product of occurrence frequency and citation-graph propagation, then evaluates the score across large publication datasets. It reports agreement with human judgments and stronger performance than alternative metrics, while noting that its analysis treats memes only as fixed character sequences.
Discussion
- The meme score is the product of occurrence frequency and propagation score, linking how often a phrase appears with how it follows citation links.Source: Discussion · p. 6
- The score can be calculated exactly and exhaustively on massive databases without arbitrary thresholds or linguistic or ontological knowledge.Source: Discussion · p. 6
- Across more than 47 million records, the meme score agreed with human annotators, outperformed alternative metrics, and its patterns were not explained by topology or timing alone.The evaluation used citation-graph randomizations, manual annotation, and comparisons with alternative metrics.Source: Discussion · pp. 6–7
- Top-ranking scientific memes show bursty temporal dynamics, which might reflect competition for scientists’ limited and fluctuating attention.Source: Discussion · p. 7
- The analysis considers only fixed character sequences, while linguistic and ontological processing could capture memes at more abstract levels.Examples include morphological variants, word cooccurrences, grammatical constructions, and rhetorical styles.Source: Discussion · p. 7
Graph randomization
The randomizations retain the original citation-network topology while assigning article texts to different nodes. In the time-preserving version, shuffling is restricted to consecutive windows of 1,000 publications.
Graph randomization
- Both randomization types preserve the original network topology but randomly reassign article texts to nodes.Source: Graph randomization · p. 7
- Time-preserving randomizations shuffle only publications within consecutive windows of 1,000 publications.Source: Graph randomization · p. 7
- After time-preserving shuffling, no publication moves more than 1,000 positions forward or backward in chronological order.Source: Graph randomization · p. 7
Human annotation
The human annotation evaluates whether phrases are meaningful physics concepts and classifies their linguistic form. Annotators independently judged a sample of meme-score terms and randomly selected comparison phrases.
Human annotation
- The annotation classified phrases as important physics concepts or entities, or as not meaningful terms or important concepts.The second classification covered noun phrases, verbs, adjectives or adverbs, and other forms.Source: Human annotation · p. 7
- The evaluation included 450 phrases: 150 top meme-score terms and two sets of 150 randomly selected phrases matched to the meme-score sample’s n-gram length distribution.One random set sampled phrases uniformly; the other weighted selection by frequency.Source: Human annotation · p. 7
- Two physics PhD students independently annotated the shuffled phrases.The sample contained the meme-score terms and both random comparison sets.Source: Human annotation · p. 7
Metrics for comparison
The comparison uses frequency- and change-based alternatives to the meme score, evaluated against a Wikipedia-derived list of physics terms. Agreement is measured by normalized area under a curve.
Metrics for comparison
- Five alternative metrics rank terms by frequency, temporal change, or frequency differences across journals.The change and journal-difference metrics each use absolute or relative values; frequency may optionally skip the first x terms.Source: Metrics for comparison · p. 7
- The alternative metrics reflect assumptions that memes are frequent but not overly common, show temporal trends, or are concentrated in particular journals.The temporal-trend rationale is attributed to [36].Source: Metrics for comparison · p. 7
- The ground-truth list contains 5,178 terms extracted from Wikipedia physics-related categories and filtered to terms appearing in at least 10 APS publications.Agreement is measured by normalized area under a curve against this list; 1 denotes perfect agreement and 0 denotes none.Source: Metrics for comparison · p. 7