Getting Started with bibnets

What bibnets Builds

Bibliometric data usually arrives as a table of papers. Each paper has fields such as authors, references, keywords, countries, institutions, source title, and year. Most bibliometric networks are projections of those fields.

The core idea is simple:

Build a sparse papers x entities matrix.
Weight the matrix with a counting method.
Multiply the matrix to obtain entity-entity or paper-paper links.
Return a standard edge list.

Internally, this is the package pipeline:

build_bipartite()
  -> apply_counting()
  -> multiply_bipartite()
  -> as_bibnets_network()

The exported builders wrap that pipeline for common bibliometric questions:

Function	Nodes	Link means
`author_network()`	authors	co-authorship, author coupling, or author co-citation
`reference_network()`	cited references	references are cited together
`document_network()`	documents	shared references, shared citers, or direct citation
`keyword_network()`	keywords	keywords appear together in papers
`source_network()`	journals/sources	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	local citation counts inside the corpus
`historiograph()`	documents	directed citation history among locally cited papers
`temporal_network()`	any builder’s nodes	the same network repeated over time windows

Every network builder returns a bibnets_network: a data frame with columns from, to, weight, and count.

count is the raw binary co-occurrence. weight is the analytical weight after counting and optional similarity normalization.

Data Used in This Vignette

The package includes small and medium example datasets:

data(biblio_data)
data(scopus_quantum_cloud)
data(open_alex_gold_open_access_learning_analytics)

small <- biblio_data
sc <- scopus_quantum_cloud
oa <- open_alex_gold_open_access_learning_analytics

nrow(small)
#> [1] 10
nrow(sc)
#> [1] 499
nrow(oa)
#> [1] 1508

biblio_data is a tiny synthetic dataset. scopus_quantum_cloud contains 499 Scopus records. open_alex_gold_open_access_learning_analytics contains 1,508 OpenAlex records with authors, countries, institutions, and primary topics.

Reading Your Own Data

For files, use read_biblio():

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

read_biblio() detects common formats from file content. You can also call a reader directly:

read_scopus("scopus.csv")
read_wos("savedrecs.txt")
read_openalex_csv("openalex_works.csv")
read_dimensions("dimensions.csv")
read_lens("lens.csv")
read_bibtex("library.bib")
read_ris("library.ris")

For a custom CSV, specify the identifier and the columns that should be split into list-columns:

data <- read_biblio(
  "custom.csv",
  format = "generic",
  id = "paper_id",
  actors = c("Authors", "Keywords"),
  sep = ";"
)

Readers return a common schema where possible:

names(sc)[1:12]
#>  [1] "id"             "title"          "year"           "journal"       
#>  [5] "doi"            "cited_by_count" "abstract"       "type"          
#>  [9] "authors"        "references"     "keywords"       "affiliations"

The most important columns for network construction are:

id: the document identifier.
authors: a list-column of author names.
references: a list-column of cited references or cited work IDs.
keywords: a list-column of author/index/topic keywords.
year: the time variable used by temporal_network().

Source-specific columns such as countries, affiliations, index_keywords, and keywords_plus are preserved when available.

Author Collaboration

The simplest author network links two authors when they appear on the same paper:

authors_full <- author_network(oa, type = "collaboration")
head(authors_full, 5)
#> # bibnets network: author_collaboration | 9 nodes · 5 edges | counting: full 
#>    from                        to                          weight  count
#> 1  MOHAMMED SAQR               SONSOLES LÓPEZ‐PERNAS        19     19
#> 2  CRISTIAN CECHINEL           ROBERTO MUÑOZ                  11     11
#> 3  DRAGAN GAŠEVIĆ            ROBERTO MARTÍNEZ‐MALDONADO      10     10
#> 4  ROBERTO MARTÍNEZ‐MALDONADO  VANESSA ECHEVERRÍA             10     10
#> 5  FERNANDA PIRES              MARCELA PESSOA                   8      8

The printed result has the standard schema:

summary(authors_full)
#> bibnets network
#> ------------------------------
#> Type       : author_collaboration
#> Counting   : full
#> Similarity : none
#> Nodes      : 4029
#> Edges      : 12270
#> Density    : 0.0015
#> Weight     : min 1  median 1  max 19
#> Top nodes  : DRAGAN GAŠEVIĆ(89), ARI KORHONEN(60), CLAUDIA SZABO(60), JUDY SHEARD(60), PAUL DENNY(56)

Use min_occur to remove very rare authors before projection:

authors_core <- author_network(oa, "collaboration", min_occur = 2)
nrow(authors_full)
#> [1] 12270
nrow(authors_core)
#> [1] 1362

Counting Methods

Counting determines how much a paper contributes to edge weights.

Full counting gives every observed co-occurrence a weight of 1:

head(author_network(small, "collaboration", counting = "full"), 5)
#> # bibnets network: author_collaboration | 5 nodes · 5 edges | counting: full 
#>    from     to       weight  count
#> 1  CHEN W   LEE K         3      3
#> 2  BROWN M  SMITH J       3      3
#> 3  BROWN M  LEE K         2      2
#> 4  JONES A  LEE K         2      2
#> 5  JONES A  SMITH J       2      2

Fractional counting reduces the influence of long author lists:

head(author_network(small, "collaboration", counting = "fractional"), 5)
#> # bibnets network: author_collaboration | 5 nodes · 5 edges | counting: fractional 
#>    from     to       weight  count
#> 1  CHEN W   LEE K         2      3
#> 2  BROWN M  SMITH J       2      3
#> 3  JONES A  SMITH J     1.5      2
#> 4  BROWN M  LEE K         1      2
#> 5  JONES A  LEE K         1      2

Harmonic counting gives more credit to earlier byline positions while keeping the paper’s total credit normalized:

head(author_network(small, "collaboration", counting = "harmonic"), 5)
#> # bibnets network: author_collaboration | 6 nodes · 5 edges | counting: harmonic 
#>    from     to       weight  count
#> 1  BROWN M  SMITH J  0.4702      3
#> 2  JONES A  SMITH J   0.371      2
#> 3  CHEN W   LEE K    0.3214      3
#> 4  CHEN W   DAVIS R  0.2222      1
#> 5  DAVIS R  JONES A  0.2222      1

First-last counting is useful only when the field’s authorship conventions make both first and last positions meaningful:

head(author_network(small, "collaboration", counting = "first_last"), 5)
#> # bibnets network: author_collaboration | 6 nodes · 5 edges | counting: first_last 
#>    from     to       weight  count
#> 1  BROWN M  SMITH J    0.49      3
#> 2  CHEN W   LEE K      0.41      3
#> 3  JONES A  SMITH J    0.33      2
#> 4  LEE K    SMITH J    0.32      2
#> 5  CHEN W   DAVIS R    0.25      1

The correct method depends on the claim being made. Use full when the question is about observed collaboration events. Use fractional when papers with many entities should not dominate. Use position-dependent methods only when author order is analytically meaningful.

Attention-Style Position Weights

The attention argument applies a smooth position profile. It is separate from counting and is available for author, keyword, country, and institution networks.

lead <- author_network(small, attention = "lead")
last <- author_network(small, attention = "last")

head(lead, 5)
#> # bibnets network: author_attention_lead | 5 nodes · 5 edges | counting: lead 
#>    from     to       weight  count
#> 1  BROWN M  SMITH J  0.3896      3
#> 2  JONES A  SMITH J  0.3437      2
#> 3  JONES A  LEE K    0.2041      2
#> 4  CHEN W   LEE K    0.2008      3
#> 5  BROWN M  CHEN W   0.1837      1
head(last, 5)
#> # bibnets network: author_attention_last | 6 nodes · 5 edges | counting: last 
#>    from     to       weight  count
#> 1  CHEN W   LEE K    0.5273      3
#> 2  BROWN M  LEE K    0.2296      2
#> 3  BROWN M  SMITH J  0.2263      3
#> 4  JONES A  LEE K    0.2041      2
#> 5  DAVIS R  SMITH J  0.1837      1

The four profiles are:

`attention`	Highest weight
`"lead"`	first position
`"last"`	last position
`"proximity"`	middle positions
`"circular"`	first and last positions

Use attention weighting when the analysis needs a transparent positional assumption rather than a named bibliometric counting convention.

Reference Co-citation

Co-citation links two cited references when they are cited together by at least one paper:

refs <- reference_network(sc, min_occur = 2)
head(refs, 5)
#> # bibnets network: reference_co_citation | 7 nodes · 5 edges | counting: full 
#>    from                            to                              weight  count
#> 1  HE K., ZHANG X., REN S., SUN …  SIMONYAN K., ZISSERMAN A., VE…      10     10
#> 2  HE K., ZHANG X., REN S., SUN …  SANDLER M., HOWARD A., ZHU M.…       8      8
#> 3  HAN S., MAO H., DALLY W.J., D…  SIMONYAN K., ZISSERMAN A., VE…       8      8
#> 4  HE K., ZHANG X., REN S., SUN …  SIMONYAN K., ZISSERMAN A., VE…       8      8
#> 5  KRIZHEVSKY A., HINTON G., LEA…  SIMONYAN K., ZISSERMAN A., VE…       7      7

Co-citation is a column-mode projection of the papers x references matrix. The nodes are references; the links come from papers that cite both references.

Similarity normalization can reduce the advantage of very frequently cited references:

refs_cos <- reference_network(sc, min_occur = 2, similarity = "cosine")
head(refs_cos, 5)
#> # bibnets network: reference_co_citation | 10 nodes · 5 edges | counting: full | similarity: cosine 
#>    from                            to                              weight  count
#> 1  ANDRI R., CAVIGELLI L., ROSSI…  CAPOTONDI A., RUSCI M., FARIS…       1      2
#> 2  BAI Y., ZENG B., LI C., ZHANG…  CASTELLI M., CLEMENTE F.M., P…       1      2
#> 3  CHEN K., ET AL., A DNN OPTIMI…  CHEN Y., ET AL., SAMBA: SINGL…       1      2
#> 4  CHELLAPILLA K., PURI S., SIMA…  CHETLUR S., WOOLLEY C., VANDE…       1      2
#> 5  ATITALLAH B.B., HU Z., BOUCHA…  CHMURSKI M., ZUBERT M., BIERZ…       1      2

Document Coupling and Citation

Bibliographic coupling links two documents when they share cited references:

coupled_docs <- document_network(sc, type = "coupling", similarity = "cosine")
head(coupled_docs, 5)
#> # bibnets network: document_coupling | 10 nodes · 5 edges | counting: full | similarity: cosine 
#>    from                to                  weight  count
#> 1  2-s2.0-85169545148  2-s2.0-85150169631  0.4671     12
#> 2  2-s2.0-85203687776  2-s2.0-85200587918  0.3872     10
#> 3  2-s2.0-85131677679  2-s2.0-85172072697  0.3424      7
#> 4  2-s2.0-85187392673  2-s2.0-85124224751   0.269     11
#> 5  2-s2.0-85161914543  2-s2.0-85100337829  0.2443     13

Direct citation is different. It keeps direction: from is the citing document and to is the cited document, but only when both documents are inside the same corpus.

direct_docs <- document_network(sc, type = "citation")
head(direct_docs, 5)
#> # bibnets network: document_citation | 0 nodes · 0 edges | counting: full

Many exported datasets cite external works that are not themselves rows in the dataset. Those external citations support co-citation and coupling, but they do not become direct-citation edges unless the cited work is also present in id.

Keyword Co-occurrence

Keyword networks are often the quickest way to inspect a corpus thematically:

kw <- keyword_network(sc, min_occur = 2)
head(kw, 5)
#> # bibnets network: keyword_co_occurrence | 5 nodes · 5 edges | counting: full 
#>    from            to              weight  count
#> 1  EDGE COMPUTING  QUANTIZATION        16     16
#> 2  DEEP LEARNING   QUANTIZATION        14     14
#> 3  DEEP LEARNING   EDGE COMPUTING      13     13
#> 4  DEEP LEARNING   FPGA                10     10
#> 5  PRUNING         QUANTIZATION        10     10

Entity labels are trimmed and uppercased during matrix construction. This means that machine learning, Machine Learning, and MACHINE LEARNING resolve to the same node.

Association strength is commonly useful for co-occurrence maps because it downweights pairs that are common only because both keywords are individually frequent:

kw_assoc <- keyword_network(sc, min_occur = 2, similarity = "association")
head(kw_assoc, 5)
#> # bibnets network: keyword_co_occurrence | 10 nodes · 5 edges | counting: full | similarity: association 
#>    from                     to                          weight  count
#> 1  AIR QUALITY PREDICTION   POST-TRAINING QUANTISATION     0.5      2
#> 2  LFSR SEED                QUANTIZATION (SIGNAL)          0.5      2
#> 3  BOOTH MULTIPLIERS        SHIFT MULTIPLIERS              0.5      2
#> 4  FERROELECTRIC CAPACITOR  SMALL-SIGNAL ANALYSIS          0.5      2
#> 5  K-NEAREST NEIGHBOR       SMART LIGHTING                 0.5      2

Countries, Institutions, and Sources

OpenAlex-style data often contains country and institution list-columns:

country_edges <- country_network(oa, counting = "fractional")
head(country_edges, 5)
#> # bibnets network: country_collaboration | 8 nodes · 5 edges | counting: fractional 
#>    from  to  weight  count
#> 1  BR    CL     9.7     11
#> 2  CA    US     9.5     13
#> 3  AU    US   8.967     15
#> 4  DE    NL   8.311     10
#> 5  CN    US     8.2     11

inst_edges <- institution_network(oa, counting = "fractional", min_occur = 2)
head(inst_edges, 5)
#> # bibnets network: institution_collaboration | 10 nodes · 5 edges | counting: fractional 
#>    from                            to                             weight  count
#> 1  FINLAND UNIVERSITY              UNIVERSITY OF EASTERN FINLAND   5.778     13
#> 2  MAASTRICHT SCHOOL OF MANAGEME…  MAASTRICHT UNIVERSITY           4.833      6
#> 3  ESCUELA SUPERIOR POLITECNICA …  MONASH UNIVERSITY               4.667      6
#> 4  UNIVERSIDADE FEDERAL DE SANTA…  UNIVERSITY OF VALPARAÍSO       4.417     10
#> 5  KUMAMOTO UNIVERSITY             KYUSHU UNIVERSITY                   4      4

Source networks use journal as the entity field. Coupling links sources that cite the same references:

source_edges <- source_network(sc, type = "coupling", min_occur = 2)
head(source_edges, 5)
#> # bibnets network: source_coupling | 5 nodes · 5 edges | counting: full 
#>    from                            to                              weight  count
#> 1  IEEE TRANSACTIONS ON CIRCUITS…  IEEE TRANSACTIONS ON COMPUTER…      48     48
#> 2  IEEE TRANSACTIONS ON CIRCUITS…  PROCEEDINGS OF THE IEEE             40     40
#> 3  IEEE TRANSACTIONS ON CIRCUITS…  IEEE TRANSACTIONS ON VERY LAR…      39     39
#> 4  IEEE JOURNAL OF SOLID-STATE C…  IEEE TRANSACTIONS ON CIRCUITS…      31     31
#> 5  IEEE TRANSACTIONS ON COMPUTER…  IEEE TRANSACTIONS ON VERY LAR…      29     29

For source, country, institution, and author coupling, min_occur is applied to the aggregated entity before building the coupling network.

Generic Co-networks

Use conetwork() when you want a projection not covered by a dedicated helper.

One-field use:

head(conetwork(sc, "keywords", min_occur = 2), 5)
#> # bibnets network: keywords_co_occurrence | 5 nodes · 5 edges | counting: full 
#>    from            to              weight  count
#> 1  EDGE COMPUTING  QUANTIZATION        16     16
#> 2  DEEP LEARNING   QUANTIZATION        14     14
#> 3  DEEP LEARNING   EDGE COMPUTING      13     13
#> 4  DEEP LEARNING   FPGA                10     10
#> 5  PRUNING         QUANTIZATION        10     10

Two-field use:

head(conetwork(sc, "authors", by = "keywords", min_occur = 2), 5)
#> # bibnets network: authors_by_keywords | 7 nodes · 5 edges | counting: full 
#>    from               to                 weight  count
#> 1  CAI H              LIU B                  36     36
#> 2  WANG Y             YIN S                  30     30
#> 3  AMROUCH H          ANAGNOSTOPOULOS I      24     24
#> 4  AMROUCH H          HENKEL J               24     24
#> 5  ANAGNOSTOPOULOS I  HENKEL J               24     24

The second example links authors through shared keywords. This is not a co-authorship network; it is a thematic-similarity network between authors.

Delimited character columns are split automatically:

toy <- data.frame(
  id = c("P1", "P2", "P3"),
  tags = c("methods; networks", "networks; R", "methods; R")
)

conetwork(toy, "tags")
#> # bibnets network: tags_co_occurrence | 3 nodes · 3 edges | counting: full 
#>    from      to        weight  count
#> 1  METHODS   NETWORKS       1      1
#> 2  METHODS   R              1      1
#> 3  NETWORKS  R              1      1

Normalization

The same raw counts can support different similarity scores:

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)
#> # bibnets network: unknown | 3 nodes · 3 edges 
#>    from            to              weight  count
#> 1  EDGE COMPUTING  QUANTIZATION        16     16
#> 2  DEEP LEARNING   QUANTIZATION        14     14
#> 3  DEEP LEARNING   EDGE COMPUTING      13     13
head(cos[, c("from", "to", "weight", "count")], 3)
#> # bibnets network: unknown | 6 nodes · 3 edges 
#>    from                    to                          weight  count
#> 1  AIR QUALITY PREDICTION  POST-TRAINING QUANTISATION       1      2
#> 2  LFSR SEED               QUANTIZATION (SIGNAL)            1      2
#> 3  BOOTH MULTIPLIERS       SHIFT MULTIPLIERS                1      2

Notice that count is unchanged. The weight column changes because normalization is applied after raw co-occurrence has been counted.

Available methods are:

normalize(to_matrix(keyword_network(small)), "cosine")
#> 23 x 23 sparse Matrix of class "dsCMatrix"
#>   [[ suppressing 23 column names 'AUTHOR NAMES', 'BIBLIOMETRICS', 'CITATION NETWORKS' ... ]]
#>                                                                      
#> AUTHOR NAMES            . . . . . . . . . . 1 . 1 . . . . . . . . . .
#> BIBLIOMETRICS           . . . . . . 1 1 . 1 . . . 1 1 1 1 1 1 . 1 . .
#> CITATION NETWORKS       . . . . 1 . . . 1 . . . . . . . . . . . . . .
#> CITATION PATTERNS       . . . . . . . . . . . 1 . . . . . . . . . . 1
#> CLUSTERING              . . 1 . . . . . 1 . . . . . . . . . . . . . .
#> CO-AUTHORSHIP           . . . . . . . . . . . . . . . . 1 . . 1 . . .
#> CO-CITATION             . 1 . . . . . . . . . . . . . . 1 . . . . . .
#> CO-OCCURRENCE           . 1 . . . . . . . . . . . . 1 . . 1 . . . 1 .
#> COMMUNITY DETECTION     . . 1 . 1 . . . . . . . . . . . . . . . . . .
#> COUPLING                . 1 . . . . . . . . . . . . . . . . 1 . . . .
#> DISAMBIGUATION          1 . . . . . . . . . . . 1 . . . . . . . . . .
#> DYNAMICS                . . . 1 . . . . . . . . . . . . . . . . . . 1
#> ENTITY RESOLUTION       1 . . . . . . . . . 1 . . . . . . . . . . . .
#> FRACTIONAL COUNTING     . 1 . . . . . . . . . . . . . . . 1 . . . . .
#> KEYWORD MAPPING         . 1 . . . . . 1 . . . . . . . . . . . . . . .
#> KNOWLEDGE DOMAINS       . 1 . . . . . . . . . . . . . . . . . . 1 . .
#> NETWORK ANALYSIS        . 1 . . . 1 1 . . . . . . . . . . . . 1 . . .
#> NORMALIZATION           . 1 . . . . . 1 . . . . . 1 . . . . . . . 1 .
#> RESEARCH FRONTS         . 1 . . . . . . . 1 . . . . . . . . . . . . .
#> SCHOLARLY COMMUNICATION . . . . . 1 . . . . . . . . . . 1 . . . . . .
#> SCIENCE MAPPING         . 1 . . . . . . . . . . . . . 1 . . . . . . .
#> SIMILARITY MEASURES     . . . . . . . 1 . . . . . . . . . 1 . . . . .
#> TEMPORAL ANALYSIS       . . . 1 . . . . . . . 1 . . . . . . . . . . .

In practice:

Use similarity = "none" for raw weighted counts.
Use similarity = "cosine" for interpretable overlap scaled by marginal size.
Use similarity = "association" when the goal is to emphasize pairs that co-occur more than expected from their individual frequencies.
Use jaccard, inclusion, or equivalence when those coefficients match a downstream method or established reporting convention.

Reducing Large Networks

Dense co-occurrence networks can be hard to inspect. bibnets provides three different reduction strategies.

edges <- author_network(oa, "collaboration")

nrow(edges)
#> [1] 12270
nrow(prune(edges, threshold = 2))
#> [1] 779
nrow(prune(edges, top_n = 5))
#> [1] 12188
nrow(filter_top(edges, n = 50))
#> [1] 956

prune(threshold = x) keeps edges with weight at least x. prune(top_n = k) keeps locally strong edges for each endpoint. filter_top(n = k) first selects the most connected nodes, then keeps edges among them.

backbone() applies the disparity filter for multiscale weighted networks:

bb <- backbone(edges, alpha = 0.05)
nrow(bb)
#> [1] 232
head(bb, 5)
#> # bibnets network: author_collaboration | 9 nodes · 5 edges | counting: full 
#>    from                        to                          weight  count
#> 1  MOHAMMED SAQR               SONSOLES LÓPEZ‐PERNAS        19     19
#> 2  CRISTIAN CECHINEL           ROBERTO MUÑOZ                  11     11
#> 3  DRAGAN GAŠEVIĆ            ROBERTO MARTÍNEZ‐MALDONADO      10     10
#> 4  ROBERTO MARTÍNEZ‐MALDONADO  VANESSA ECHEVERRÍA             10     10
#> 5  FERNANDA PIRES              MARCELA PESSOA                   8      8

The disparity filter asks whether an edge is unusually strong relative to at least one endpoint’s local strength distribution. This is different from a global weight cutoff and can preserve meaningful edges attached to smaller nodes.

Temporal Networks

temporal_network() runs any network builder over time windows:

tn <- temporal_network(oa, author_network, "collaboration", window = 3)
names(tn)
#> [1] "2011-2013" "2014-2016" "2017-2019" "2020-2022" "2023-2025" "2026-2026"

Fixed windows are non-overlapping. Sliding windows overlap:

tn_slide <- temporal_network(
  oa,
  author_network,
  "collaboration",
  window = 3,
  step = 1,
  strategy = "sliding"
)

names(tn_slide)
#>  [1] "2011-2013" "2012-2014" "2013-2015" "2014-2016" "2015-2017" "2016-2018"
#>  [7] "2017-2019" "2018-2020" "2019-2021" "2020-2022" "2021-2023" "2022-2024"
#> [13] "2023-2025" "2024-2026"

Cumulative windows always start at the first observed year and grow forward:

tn_cum <- temporal_network(
  oa,
  author_network,
  "collaboration",
  window = 3,
  strategy = "cumulative"
)

names(tn_cum)
#>  [1] "2011-2013" "2011-2014" "2011-2015" "2011-2016" "2011-2017" "2011-2018"
#>  [7] "2011-2019" "2011-2020" "2011-2021" "2011-2022" "2011-2023" "2011-2024"
#> [13] "2011-2025" "2011-2026"

Each returned edge list has a window column. Windows with fewer than two records or no surviving edges are omitted. If a builder errors inside a window, temporal_network() reports a warning with the window label.

Local Citations and Historiographs

local_citations() counts how often each document is cited by other documents inside the same dataset:

lcs <- local_citations(sc)
head(lcs, 5)
#>                    id lcs gcs year
#> 1 2-s2.0-105007159281   0   0 2025
#> 2 2-s2.0-105006878874   0   0 2025
#> 3  2-s2.0-85211114952   0   0 2024
#> 4 2-s2.0-105001072133   0   0 2025
#> 5  2-s2.0-85210832535   0   5 2025
#>                                                                                                                                  title
#> 1                                          Quantum Computing in the RAN with Qu4Fec: Closing Gaps Towards Quantum-based FEC Processors
#> 2                                                An FPGA-based bit-level weight sparsity and mixed-bit accelerator for neural networks
#> 3                             FQP: A Fibonacci Quantization Processor with Multiplication-Free Computing and Topological-Order Routing
#> 4                                                   SysCIM: A Heterogeneous Chip Architecture for High-Efficiency CNN Training at Edge
#> 5 Integer-Valued Training and Spike-Driven Inference Spiking Neural Network for High-Performance and Energy-Efficient Object Detection
#>                                                                                                                                 journal
#> 1                                                               Proceedings of the ACM on Measurement and Analysis of Computing Systems
#> 2                                                                                                       Journal of Systems Architecture
#> 3                                                                                            Proceedings - Design Automation Conference
#> 4                                                                      IEEE Transactions on Very Large Scale Integration (VLSI) Systems
#> 5 Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 
#>                            doi
#> 1              10.1145/3727128
#> 2 10.1016/j.sysarc.2025.103463
#> 3      10.1145/3649329.3656502
#> 4   10.1109/TVLSI.2025.3526363
#> 5 10.1007/978-3-031-73411-3_15

historiograph() builds a directed citation graph among the top locally cited documents:

h <- historiograph(sc, n = 10)
h$nodes
#> [1] id      lcs     gcs     year    title   journal doi    
#> <0 rows> (or 0-length row.names)
head(h$edges, 5)
#> [1] from      to        year_from year_to  
#> <0 rows> (or 0-length row.names)

This requires reference strings or IDs to match document IDs in the same data frame. If the cited works are external to the corpus, local citation counts will be low or zero even when global citation counts are high.

Exporting Results

The default edge list is already useful for many tools:

edges <- keyword_network(sc, min_occur = 2)
head(edges, 5)
#> # bibnets network: keyword_co_occurrence | 5 nodes · 5 edges | counting: full 
#>    from            to              weight  count
#> 1  EDGE COMPUTING  QUANTIZATION        16     16
#> 2  DEEP LEARNING   QUANTIZATION        14     14
#> 3  DEEP LEARNING   EDGE COMPUTING      13     13
#> 4  DEEP LEARNING   FPGA                10     10
#> 5  PRUNING         QUANTIZATION        10     10

Convert to a sparse matrix:

m <- to_matrix(edges)
m[1:4, 1:4]
#> 4 x 4 sparse Matrix of class "dgCMatrix"
#>                ACCELERATION ACCELERATOR ACCURACY AI ACCELERATOR
#> ACCELERATION              .           .        .              .
#> ACCELERATOR               .           .        .              .
#> ACCURACY                  .           .        .              .
#> AI ACCELERATOR            .           .        .              .

Prepare Gephi tables:

gephi <- to_gephi(edges)
head(gephi$nodes, 3)
#>               Id          Label
#> 1 EDGE COMPUTING EDGE COMPUTING
#> 2  DEEP LEARNING  DEEP LEARNING
#> 3        PRUNING        PRUNING
head(gephi$edges, 3)
#>           Source         Target Weight       Type count
#> 1 EDGE COMPUTING   QUANTIZATION     16 Undirected    16
#> 2  DEEP LEARNING   QUANTIZATION     14 Undirected    14
#> 3  DEEP LEARNING EDGE COMPUTING     13 Undirected    13

Write GraphML without adding an XML dependency:

xml <- to_graphml(edges)
cat(substr(xml, 1, 300))
#> <?xml version="1.0" encoding="UTF-8"?>
#> <graphml xmlns="http://graphml.graphdrawing.org/graphml"
#>          xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
#>          xsi:schemaLocation="http://graphml.graphdrawing.org/graphml
#>          http://graphml.graphdrawing.org/graphml/1.0/graphml.xsd">
#>   <ke

Optional graph objects are available when the suggested packages are installed:

if (requireNamespace("igraph", quietly = TRUE)) {
  g <- to_igraph(edges)
}

if (requireNamespace("tidygraph", quietly = TRUE)) {
  tg <- to_tbl_graph(edges)
}

if (requireNamespace("cograph", quietly = TRUE)) {
  cg <- to_cograph(edges)
}

Interpreting a `bibnets_network`

The object stores construction metadata as attributes:

edges <- author_network(oa, "collaboration", counting = "harmonic")

attr(edges, "network_type")
#> [1] "author_collaboration"
attr(edges, "counting")
#> [1] "harmonic"
attr(edges, "similarity")
#> [1] "none"

The print() method reports the network type, node count, edge count, counting method, and similarity method. summary() reports basic network and weight summaries:

summary(edges)
#> bibnets network
#> ------------------------------
#> Type       : author_collaboration
#> Counting   : harmonic
#> Similarity : none
#> Nodes      : 4029
#> Edges      : 12270
#> Density    : 0.0015
#> Weight     : min 1.43e-05  median 0.0106  max 1.41
#> Top nodes  : DRAGAN GAŠEVIĆ(89), ARI KORHONEN(60), CLAUDIA SZABO(60), JUDY SHEARD(60), PAUL DENNY(56)

These attributes are meant to make downstream output easier to audit. A saved edge list should still say how it was produced.