Getting Started with bibnets

In bibnets: Importing, Constructing, and Exporting Bibliometric Networks

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(bibnets)

Introduction

bibnets constructs bibliometric networks from scholarly metadata. It imports the export formats of the major bibliographic databases, converts them internally to a common tabular representation, and projects that representation into networks through a single function per network type. The package covers the standard constructions and adds several that are less commonly available: position-based attention weighting, aggregation of entities into higher-level networks, a range of counting and similarity weights, and temporal construction over time windows.

Data import

bibnets reads Scopus, Web of Science, OpenAlex, Lens.org, Dimensions, Crossref, BibTeX, and RIS exports. read_biblio() detects the format from file content and dispatches to the corresponding reader; all readers return an identical schema, so records from different databases can be combined without manual reconciliation. Multi-valued fields — authors, references, and keywords — are parsed into list-columns. A data frame already in this form is used directly, by naming the relevant column, without a reader.

Network builders

Dedicated builders construct co-authorship, co-citation, bibliographic coupling, keyword co-occurrence, direct citation, and historiograph networks; a generic builder covers other projections. The builders share one interface and return the same edge list, so the network type is determined by the function name.

Weighting and aggregation

Counting methods determine each publication's contribution to an edge. They range from full and fractional counting to position-aware schemes (harmonic, geometric, golden-ratio, first, last, and first–last), and six similarity measures — cosine, association strength, Jaccard, inclusion, and equivalence — rescale the projected weights.

Attention weighting assigns each author a positional weight that sums to one across the byline, so a publication's credit is distributed by byline position rather than equally and a large author list does not dominate the network. Aggregation pools the references or members of a group to construct collaboration and coupling networks among countries, institutions, or sources rather than individuals. Temporal construction applies any builder over fixed, sliding, or cumulative windows, and a disparity-filter backbone retains the edges that are significant relative to each node's strength.

Implementation

The incidence matrix is stored as a sparse dgCMatrix and projected with crossprod() or tcrossprod(); edges are extracted without forming a dense node-by-node matrix, so memory scales with the number of non-zero co-occurrences rather than with the square of the vocabulary. The package imports only Matrix, stats, and utils.

Export formats

Constructed networks are exported to igraph, tidygraph, cograph, Gephi, GraphML, and sparse-matrix representations.

Output

Every builder returns a bibnets_network: a tidy data frame with four columns —

• from, to — the two endpoints of an edge,
• count — the raw binary co-occurrence count for that pair,
• weight — the analytical weight after counting and optional similarity normalization.

With counting = "full" and similarity = "none", weight equals count. They diverge once fractional counting or a similarity measure is applied.

The builders, at a glance:

| Function | Nodes | An edge means | |---|---|---| | author_network() | authors | co-authorship, author coupling, or co-citation | | reference_network() | cited references | two references cited together | | document_network() | documents | shared references, shared citers, or direct citation | | keyword_network() | keywords | two keywords appear together | | source_network() | journals | sources share references or are co-cited | | country_network() | countries | countries collaborate or share references | | institution_network() | institutions | institutions collaborate or share references | | conetwork() | any field | entities co-occur, or share values of another field | | local_citations() | documents | within-corpus citation counts | | historiograph() | documents | directed citation history among top-cited papers | | temporal_network() | any builder's nodes | the same network over time windows |

Quick start

You do not need a special reader. Any data frame with one row per paper works — point a builder at the column that holds the entity and tell it the delimiter:

papers <- data.frame(
  `Author Names` = c("Smith J, Doe A, Lee K", "Smith J, Lee K",
                     "Doe A, Lee K", "Smith J, Doe A"),
  check.names = FALSE
)

author_network(papers, authors = "Author Names", sep = ",")

If your data is a scholarly export instead, read it first — the format is detected from the file content — then build with the defaults:

data    <- read_biblio("scopus.csv")
authors <- author_network(data, type = "collaboration")

Either way the result is the same four-column edge list, ready to inspect, prune, or export.

Reading your own data

Scholarly exports

read_biblio() accepts a file, a folder, or several files, and detects Scopus, Web of Science, OpenAlex, BibTeX, RIS, Lens.org, and Dimensions from the content:

data <- read_biblio("export.csv")
data <- read_biblio("folder_with_exports/")
data <- read_biblio(c("part_1.csv", "part_2.csv"))

The format-specific readers can also be called directly (read_scopus(), read_wos(), read_openalex_csv(), read_dimensions(), read_lens(), read_bibtex(), read_ris()).

A custom CSV

For a CSV that matches no known export, map each source column onto a standard field by name — authors, keywords, references, countries, affiliations, or journal. Naming any of them reads the file as a generic CSV, so you do not pass format yourself:

data <- read_biblio(
  "custom.csv",
  id       = "paper_id",
  authors  = "Author Names",
  keywords = "Tags",
  sep      = ","
)

Each mapped column is split on sep into the standard list-column, so afterwards every builder works with its defaults.

A plain data frame, directly

As the quick start showed, you can skip the reader entirely and let the builder split a column for you. The same column arguments are available on every builder:

author_network(my_df, authors = "Author Names", sep = ",")
keyword_network(my_df, keywords = "Tags",       sep = ",")

The work identifier is the id column. You need not supply one: when no id column is present each row is treated as one document; pass id = "paper_id" to use a differently-named column. Surrounding quotes are stripped by default (strip_quotes = TRUE), and in a coupling network the references column takes its own references_sep. The companion vignette("reading-data") covers every reader and these options in full.

The standard schema

Readers return a common set of columns:

data(scopus_quantum_cloud)
sc <- scopus_quantum_cloud
names(sc)[1:12]

The columns that matter for network construction are id, the list-columns authors / references / keywords, and year (used by temporal_network()). Source-specific extras such as countries, affiliations, and keywords_plus are kept when available.

Datasets used here

data(biblio_data)
data(learning_analytics)

small <- biblio_data            # tiny, synthetic
oa    <- learning_analytics     # 1,508 OpenAlex records on learning analytics

c(small = nrow(small), scopus = nrow(sc), openalex = nrow(oa))

Author collaboration

Two authors are linked when they appear on the same paper:

authors <- author_network(oa, type = "collaboration")
head(authors, 5)
summary(authors)

Use min_occur to drop rare authors before projection:

nrow(author_network(oa, type = "collaboration"))
nrow(author_network(oa, type = "collaboration", min_occur = 2))

Counting methods

counting controls how much each paper contributes to an edge:

head(author_network(small, type = "collaboration", counting = "full"), 3)
head(author_network(small, type = "collaboration", counting = "fractional"), 3)
head(author_network(small, type = "collaboration", counting = "harmonic"), 3)
head(author_network(small, type = "collaboration", counting = "first_last"), 3)

The available methods differ in how they weight the rows or positions before projection:

| Method | What it does | Trade-off | When to use | |---|---|---|---| | "full" | Leaves the binary incidence matrix unchanged; for positional author weights, every listed entity receives weight 1. | Large teams or long lists create many full-strength pairs. | Use for raw event counts where every observed co-occurrence should count equally. | | "fractional" | For symmetric networks, each row contributes 1 / (n - 1) to pairs when n > 1; for coupling it uses 1 / n; positional use gives each entity 1 / n. | Reduces large-list dominance but treats all positions equally. | Use when each paper or reference list should have limited influence and position is not meaningful. | | "paper" | For symmetric networks, each paper's pair budget is scaled by 2 / (n * (n - 1)); for coupling it uses 1 / n. | Normalizes at the paper level, so very large and very small papers can contribute comparable total pair mass. | Use when publications, rather than individual author/entity pairs, should be the main unit of contribution. | | "strength" | Multiplies entity columns by the square root of inverse document frequency, sqrt(log(n_works / entity_frequency)); row-size scaling for coupling is deferred to projection. | Downweights ubiquitous entities and emphasizes rarer shared entities; values are less like direct counts. | Use for coupling or profile similarity where common references, keywords, or entities should carry less evidence. | | "harmonic" | Uses positional weights proportional to 1 / position, normalized to sum to one. | Strongly favors early positions while still giving every later position some credit. | Use when author order matters and early authorship should dominate without excluding later authors. | | "arithmetic" | Uses a linear decline from first to last, proportional to n - position + 1, normalized. | Gives a gentler first-author advantage than geometric methods. | Use when byline order matters but credit should decrease steadily rather than sharply. | | "geometric" | Uses weights proportional to 0.5^(position - 1), normalized. | Concentrates credit heavily at the front of the byline. | Use when the first few positions are expected to carry most of the contribution. | | "adaptive_geometric" | Uses a geometric sequence normalized so the first-to-last weight ratio equals n (2/3, 1/3 for two authors). | Adapts the steepness to team size, making long bylines more front-loaded. | Use when first-author emphasis should increase with the number of authors. | | "golden" | Uses golden-ratio decay, proportional to phi^(n - position), normalized. | More front-loaded than arithmetic but less abrupt than fixed halving. | Use as a moderate positional decay when author order matters but geometric halving is too strong. | | "first" | Gives weight 1 to the first position and 0 to all others. | Ignores all non-first contributors. | Use for strict first-author analyses. | | "last" | Gives weight 1 to the last position and 0 to all others. | Ignores all non-last contributors. | Use where last authorship represents the analytical role of interest, such as senior or PI credit. | | "first_last" | With two authors, assigns 0.5 and 0.5; otherwise gives first and last authors an elevated weight and middle authors a baseline weight, all normalized. | Highlights both endpoints while still retaining middle-author credit. | Use in fields where first and last positions have distinct credit or leadership meanings. | | "position_weighted" | Uses the supplied position_weights vector, extending the last value to longer bylines, then normalizes. | Puts the burden of choosing defensible weights on the analyst. | Use when you have field-specific or study-specific positional weights. |

Attention weights

Standard bibliometric co-authorship networks treat every byline position as equivalent: a first author who conceived and drove the work is weighted identically to a fifteenth contributor who provided a single instrument reading. On hyper-authored papers this produces dense, low-meaning co-authorship ties that drown out the focused two- or three-author collaborations that often signal the sharpest intellectual kinship. The attention weighting feature in bibnets is designed to correct this. The name is an honest analogy to the attention mechanism in large language models: just as a transformer assigns a normalized probability distribution across the tokens in a sequence — concentrating weight on what matters, spreading it thin over the rest — bibnets assigns each author on a paper a positional weight that sums to one across the full byline. A fifty-author paper therefore contributes exactly one unit of connection budget, the same as a two-author paper, and the distribution of that budget reflects authorship conventions: "lead" concentrates weight on the first author, "last" on the senior or PI position, "proximity" rewards the central authors, and "circular" rewards both ends jointly. The weights are a fixed positional prior, not learned content-based attention, but they carry real scholarly meaning, and activating them requires nothing more than passing attention = "lead" (or any of the three alternatives) to any of the author, keyword, country, or institution network functions.

| attention | Weight vector | Scholarly assumption | When it fits | |---|---|---|---| | "lead" | Quadratic drop from the first position: the first position has raw weight n^2, inner positions decline as the byline advances, and the last position has raw weight 1, then all weights are normalized. | The lead author is the main intellectual driver. | Use in first-author-oriented fields or questions about lead contribution. | | "last" | Quadratic rise to the last position: the first position has raw weight 1, inner positions rise across the byline, and the last position has raw weight n^2, then all weights are normalized. | The last author represents senior, supervisory, or PI contribution. | Use in disciplines where last authorship marks lab leadership or supervision. | | "proximity" | Pyramid profile using min(position, n + 1 - position): first and last positions have raw weight 1, inner positions increase toward the middle, and central positions are highest. | Central byline positions deserve the most attention. | Use when the question treats middle-position contributors as the focal group. | | "circular" | Edge profile using max(position, n + 1 - position): first and last positions have the largest raw weight, while inner positions decline toward the center. | Both ends of the byline are prominent. | Use where lead and senior positions jointly matter more than middle positions. |

attention applies a smooth positional profile instead of a named counting scheme (available for author, keyword, country, and institution networks):

head(author_network(small, attention = "lead"), 3)

Reference co-citation

Two references are linked when a paper cites both:

refs <- reference_network(sc, min_occur = 2)
head(refs, 5)

A similarity measure offsets the advantage of very frequently cited works:

head(reference_network(sc, min_occur = 2, similarity = "cosine"), 3)

Document coupling and citation

Coupling links two documents that share cited references:

head(document_network(sc, type = "coupling", similarity = "cosine"), 5)

Direct citation is directed — from cites to — and only within the corpus (the cited work must also be a row in the data):

head(document_network(sc, type = "citation"), 5)

Keyword co-occurrence

kw <- keyword_network(sc, min_occur = 2)
head(kw, 5)

Labels are trimmed and upper-cased during construction, so machine learning, Machine Learning, and MACHINE LEARNING are one node. Association strength is a common choice for co-occurrence maps:

head(keyword_network(sc, min_occur = 2, similarity = "association"), 3)

Countries, institutions, and sources

head(country_network(oa, counting = "fractional"), 5)
head(institution_network(oa, counting = "fractional", min_occur = 2), 5)
head(source_network(sc, type = "coupling", min_occur = 2), 5)

For coupling networks, min_occur is applied to the aggregated entity before the network is built.

Generic co-networks

conetwork() covers projections without a dedicated wrapper. One field links entities that co-occur; a second field (by) links them through a shared value:

head(conetwork(sc, "keywords", min_occur = 2), 3)
head(conetwork(sc, "authors", by = "keywords", min_occur = 2), 3)

The second result links authors through shared keywords — a thematic similarity network, not a co-authorship one.

Normalization

The same raw counts support different similarity scores; only weight changes, count does not:

none <- keyword_network(sc, min_occur = 2, similarity = "none")
cos  <- keyword_network(sc, min_occur = 2, similarity = "cosine")
head(none[, c("from", "to", "weight", "count")], 3)
head(cos[,  c("from", "to", "weight", "count")], 3)

normalize() uses the diagonal of the projected matrix as each node's total occurrence count:

| Similarity | Denominator | Meaning | When to use | |---|---|---|---| | "none" | No denominator; the projected matrix is returned as raw weighted co-occurrence, with the diagonal removed by the network builder unless self-loops are requested. | weight stays on the same scale as the counted projection. | Use when absolute co-occurrence or counted edge strength is the quantity of interest. | | "cosine" | Square root of the product of the two node totals. | Symmetric size correction; pairs are high when their overlap is large relative to both nodes' frequencies. | Use as a general-purpose correction for very frequent nodes while preserving a familiar similarity scale. | | "association" | Product of the two node totals. | Symmetric association-strength normalization; strongly penalizes pairs involving very frequent nodes. | Use for co-occurrence maps where you want rare, unexpectedly tight pairings to stand out. | | "jaccard" | Sum of the two node totals minus their observed edge value. | Symmetric overlap over a union-like total. | Use when the edge should represent shared occurrence as a share of either node's combined footprint. | | "inclusion" | The smaller of the two node totals. | Symmetric containment-oriented score; it reaches high values when the smaller node mostly appears with the larger one. | Use when subset or specialization relationships are more important than balanced overlap. | | "equivalence" | Product of the two node totals, with the edge value squared before division. | Cosine-like normalization with stronger penalty for weak or occasional overlap. | Use when following equivalence-index conventions or when only consistently paired nodes should remain strong. |

Reducing large networks

edges <- author_network(oa, type = "collaboration")
c(all        = nrow(edges),
  threshold  = nrow(prune(edges, threshold = 2)),
  top_n      = nrow(prune(edges, top_n = 5)),
  top_nodes  = nrow(filter_top(edges, n = 50)))

• prune(threshold = x) — absolute edge-weight cutoff.
• prune(top_n = k) — keep each node's strongest edges.
• filter_top(n = k) — keep edges among the most-connected nodes.

backbone() applies the disparity filter, which keeps edges that are strong relative to a node's local strength distribution — not a global cutoff:

bb <- backbone(edges, alpha = 0.05)
nrow(bb)

Temporal networks

temporal_network() runs any builder over time windows (fixed, sliding, or cumulative):

tn <- temporal_network(oa, author_network, "collaboration", window = 3)
names(tn)

Each window's edge list carries a window column. Windows with fewer than two records, or no surviving edges, are dropped; a builder error inside a window becomes a warning labelled with that window.

Local citations and historiographs

local_citations() counts how often each document is cited by others in the same corpus; historiograph() builds the directed citation graph among the top-cited documents:

head(local_citations(sc), 5)

h <- historiograph(sc, n = 10)
h$nodes
head(h$edges, 5)

Both require reference strings or IDs to match document IDs in the data; if the cited works are external, local counts stay low.

Author-name normalization

parse_names() reorders and splits author names (it recognizes Last, First, SURNAME Initials, and First Last). Because node identity is fixed when a network is built, normalize before building so that two spellings of one author merge:

parse_names(c("Saqr, Mohammed", "WANG Y", "Mohammed Saqr"))

See vignette("parsing-author-names") for the full treatment.

Exporting

The edge list is already usable; converters cover the common targets:

edges <- keyword_network(sc, min_occur = 2)

m <- to_matrix(edges)            # sparse adjacency matrix
m[1:4, 1:4]

gephi <- to_gephi(edges)         # Gephi node/edge tables
head(gephi$edges, 3)

cat(substr(to_graphml(edges), 1, 200))   # GraphML, no XML dependency

to_igraph(), to_tbl_graph(), and to_cograph() are available when their (suggested) packages are installed.

Reading a bibnets_network

The object records how it was built, as attributes:

edges <- author_network(oa, type = "collaboration", counting = "harmonic")
c(type     = attr(edges, "network_type"),
  counting = attr(edges, "counting"),
  sim      = attr(edges, "similarity"))

summary(edges)

print() reports the network type, node and edge counts, and the counting and similarity methods — so a saved edge list always says how it was made.

References

The methodology implemented in bibnets is described in:

López-Pernas, S., Saqr, M., & Apiola, M. (2023). Scientometrics: A Concise Introduction and a Detailed Methodology for Mapping the Scientific Field of Computing Education Research. In M. Apiola, S. López-Pernas, & M. Saqr (Eds.), Past, Present and Future of Computing Education Research: A Global Perspective (pp. 79–99). Springer Nature Switzerland AG. https://doi.org/10.1007/978-3-031-25336-2_5

Saqr, M., López-Pernas, S., Conde, M. Á., & Hernández-García, Á. (2024). Social Network Analysis: A primer, a guide and a tutorial in R. In M. Saqr & S. López-Pernas (Eds.), Learning Analytics Methods and Tutorials: A Practical Guide Using R (pp. 491–518). Springer, Cham. https://doi.org/10.1007/978-3-031-54464-4_15

bibnets documentation built on June 19, 2026, 1:06 a.m.