Skip to contents

Changelog

Source: NEWS.md

networkflow (development version)

  • No change yet since 0.1.0.

networkflow 0.1.0 (first CRAN release)

Earlier development steps are numbered 0.0.9.900x, from the first step (rewrite) to the latest; they are listed from the most recent to the oldest.

networkflow 0.0.9.9001 (2026-11-01 to 2026-11-05)

Second step: the pipeline goes beyond bibliographic coupling, networks are built with one explicit method, time windows and dynamic clusters are reworked, the package is lighter and faster, and it was checked on samples of the Scopus and Web of Science databases of the authors. The changes are given with respect to 0.0.9.9000.

Breaking changes

  • build_network() and build_dynamic_networks(): one required argument, method, replaces projection_method, cooccurrence_method and model. It is either one of the five similarity measures of biblionetwork ("angle", "strength", "similarity", "jaccard", "vladutz") or one of the null models of the statistical backbones of the backbone package ("sdsm", "fdsm", "fixedfill", "fixedrow", "fixedcol", with alpha). There is no default. The former arguments give an error that shows the translation (projection_method = "structured", cooccurrence_method = "coupling_angle" becomes method = "angle").
  • build_dynamic_networks() needs time_variable and time_window, and points to build_network() otherwise (it used to build the single network silently).
  • overlapping_window is replaced by window_step, the number of units between the starts of two windows (window_step = 1 replaces overlapping_window = TRUE). The arguments are reordered, and ... only accepts the arguments of filter_components(). Windows of a single time unit are named after the unit ("1990" instead of "1990-1990").
  • merge_dynamic_clusters() and merge_diagnostics() have a required method: "majority" (the former double majority) or "jaccard" (new). threshold_similarity has no default any more: calls of earlier versions need method = "majority" and a threshold.
  • name_clusters(): name_merged_clusters is renamed name_dynamic_clusters (the old name still works, with a deprecation warning).
  • Argument names are harmonised: cluster_id replaces module_col (add_node_roles()) and intertemporal_cluster_column (networks_to_alluv(), minimize_crossing_alluvial(), plot_alluvial()), and graphs replaces list_graph (merge_dynamic_clusters(), merge_diagnostics()). add_node_roles() takes weights, as add_clusters() and tidygraph (NULL uses the weight edge column if any), and writes the roles in a column named roles (it was role_ga). extract_tfidf() keeps data, as it also takes plain tables.
  • plot_networks() returns a named list of ggplot objects for a list of graphs.
  • networks_to_alluv() returns window as a factor in the order of the list of graphs.
  • biblionetwork >= 0.2.0 is required.

New features

  • Beyond bibliographic coupling: co-citation (projection_on = "target"), time carried by the links (time_in = "edges"), coupling of entities that publish documents, such as authors or journals (document_id, through biblionetwork::entity_network()), and a compute_size that counts the citing documents of each node.
  • Dates (Date, POSIXct, ISO dates) and finer time units (time_unit: year, quarter, month, week, day).
  • With verbose = TRUE, the building functions say which method they apply, what it measures and where it comes from.
  • merge_dynamic_clusters(method = "jaccard"): overlap matching with a Jaccard threshold (Greene et al., 2010), each past identifier going to one cluster at most.
  • merge_diagnostics() and plot_merge_diagnostics() show how the dynamic clusters depend on the rule and the threshold.
  • layout_networks() has two new layouts, "forceatlas2" (Jacomy et al.,
    1. and "umap" (McInnes et al., 2018).
  • minimize_crossing_alluvial() reduces the crossings of the flows themselves (a local search, or the barycentre heuristic for large alluvials) and reports them.
  • plot_alluvial() accepts the list of graphs directly, and keeps the main clusters with min_share_cluster and min_windows_cluster.
  • name_clusters() with tf-idf skips the terms that overlap a better one (“business” after “business cycle”).
  • Windows without any edge give an empty network that the following steps accept; nodes or links with a missing time or identifier are dropped with a warning; a target linked to more than half of the sources (such as a code for unmatched references) is reported.

Speed and memory

Measured on samples of 30,000 documents of Scopus and of the Web of Science:

  • build_dynamic_networks() is three times faster (24 s instead of 83 s for 11 windows of the Web of Science sample).
  • merge_diagnostics() is about seven times faster (8 s instead of 63 s).
  • minimize_crossing_alluvial() takes about a second instead of 5 to 35 minutes on alluvials of 250 to 550 clusters, with bounded memory.
  • On the whole pipeline, the peak of memory falls from 13.9 GB to 3.9 GB on the Web of Science sample.
  • plot_networks() draws straight edges on graphs of more than max_arcs edges (20,000): 2 seconds and 1 GB instead of 100 seconds and 11 GB for 140,000 edges.
  • UMAP uses the nearest neighbours instead of all the distances (1.4 GB instead of 5.7 GB on 9,600 nodes), and ForceAtlas2 computes its repulsion by blocks.
  • 16 packages are imported instead of 22: forcats, stringr and stopwords are removed, and ggalluvial, tidytext and lifecycle are suggested. The functions that need a suggested package ask to install it.

Bug fixes

  • plot_networks() failed on PDF devices where the font of ggraph::theme_graph() (Arial Narrow) is not installed.
  • minimize_crossing_alluvial() sorted the clusters on a string of component numbers, wrong as soon as there were ten components or more.
  • Invalid edges_threshold (0, a string) failed deep inside biblionetwork; it is now checked first.

Documentation and website

  • Five guides on the website: “From references to networks”, “Finding and reading clusters”, “Drawing networks”, “Following clusters over time” and “Beyond bibliographic coupling”, with the methods, their formulas and the literature behind them.
  • The documentation of each function explains its method and its choices, with formulas displayed on the website, the origin of each measure, and the cost of the slow methods (statistical backbones, ForceAtlas2, Kamada-Kawai).

networkflow 0.0.9.9000 (2026-11-01)

First step: rewrite of the documentation and of the code of the former GitHub version (0.0.9), with tests and benchmarks. To be released on CRAN as 0.1.0.

Breaking changes

  • merge_dynamic_clusters(): similarity_type must be given ("complete" or "partial"). Its former default could not be used anyway (it failed with “the condition has length > 1”).
  • add_clusters(): cluster identifiers are numbered by decreasing size for every algorithm (“01” is the largest cluster). Leiden identifiers were previously in the order returned by igraph: partitions are unchanged, but labels may differ from those obtained with an earlier version.
  • add_clusters(): Louvain, fast greedy, walktrap and Infomap now use the weight edge column, as documented. Through tidygraph, they used to ignore it (only Leiden used the weights): partitions obtained with these algorithms on weighted networks change.
  • build_network(): model, alpha and backbone_args are explicit arguments; ... only goes to filter_components().
  • build_dynamic_networks() keeps node_size computed by compute_size instead of duplicating an existing node_size column.

Removed

Other changes

  • add_clusters(): the default of n_iterations is now -1 (Leiden iterates until the partition is stable) instead of 1000. Results are close but not identical to those obtained with 1000 iterations, and much faster on large networks.
  • launch_network_app(): the UI, the server and the plot are now separate internal functions, tested with shiny::testServer(); the default layout is "fr" (Kamada-Kawai reads similarity weights as distances).
  • citation("networkflow") works (inst/CITATION).
  • Credits: Alexandre Truc initiated the alluvial method; Thomas Delcey and Aurélien Goutsmedt maintain the package.
  • GitHub Actions: an R-CMD-check workflow (three systems) was added, and the pkgdown workflow installs Quarto.

Bug fixes

  • add_node_roles() computed the within-module degree and the participation coefficient wrongly for the node at the to end of each edge: all its ties were counted as internal to its cluster. Roles computed with earlier versions should be recomputed.
  • add_clusters(seed = ) no longer changes the random-number stream of the session: the seed is set locally with withr::local_seed() (CRAN policy).
  • add_clusters() returns an error instead of NULL for an input that is neither a tbl_graph nor a list of them.
  • add_clusters() and merge_dynamic_clusters() return graphs with the node table active (they left the edge table active).
  • merge_dynamic_clusters() can be run again on graphs that already have dynamic clusters, works on unnamed lists and on graphs whose edges lack the *_from/*_to cluster columns, and warns when consecutive windows share no node (successive windows of documents cannot be matched).
  • build_network(projection_method = "statistical", filter_components = TRUE) failed with “unused arguments”.
  • build_dynamic_networks() no longer fails on a window where no target is shared by two sources (it returns a graph without edges).
  • name_clusters(method = "given_column", name_merged_clusters = TRUE) failed; the check that tidygraph was attached (through sessionInfo()) is removed.
  • prepare_label_networks() ignored x and y for a single graph.
  • plot_networks(print_plot_code = TRUE) used a wrong title for lists.
  • filter_components() and layout_networks() returned their result invisibly.
  • layout_networks() no longer triggers ggraph’s warning about existing x and y columns.
  • extract_tfidf() no longer triggers data.table’s “shallow copy” warning, and missing texts no longer produce the term “na”.
  • color_networks(): with more than 14 categories, the default palette used 13 colours; it now uses 14 and warns that colours are recycled.

networkflow 0.0.9

Development version, distributed from GitHub (it was numbered 1.0.0).

Deprecated and new functions:

networkflow 0.0.0.9000

Launch of a development version