Skip to contents

Build the soft cumulative-clustering trajectory DAG

Source: R/soft_trajectory_dag.R
sniff_trajectory_dag.Rd

Reconstructs, from docs_per_group alone, the full temporal DAG of cluster-year nodes: vertices are every cluster whose distinct-document count is at least min_group_size (isolated nodes included), edges are consecutive-year Jaccard links (top k_out per source, weight >= jaccard_min). Each node is labelled with the final group its heaviest-successor walk reaches (terminal_group) and whether it is a birth (no strong predecessor). This DAG is the first-class object the soft trajectories and group-formation views are read from.

Usage

sniff_trajectory_dag(
  x,
  min_group_size = 10,
  jaccard_min = 0.05,
  k_out = 2,
  similarity = c("overlap", "coupling"),
  references = NULL
)

Arguments

x

The membership tibble docs_per_group (group_id, document_id, network_until, group) or a sniff_groups_lineage() object (its $docs_per_group is used).

min_group_size

Minimum distinct-document count for a node (default 10).

jaccard_min

Minimum Jaccard weight to keep an edge (default 0.05).

k_out

Max outgoing edges kept per source node (default 2).

similarity

How to weight consecutive-year edges: "overlap" (default) uses the Jaccard of the nodes' shared documents (the cumulative carry-over); "coupling" uses the Jaccard of their shared cited references (bibliographic coupling) and requires references. Both detectors (sniff_trajectory_braid(), sniff_trajectory_channel()) route either DAG unchanged.

references

When similarity = "coupling", a data frame with columns document_id and feature (one cited reference per row), the same shape as a content element of sniff_trajectory_coherence(). Ignored for "overlap".

Value

A birddog_dag object: a classed list with graph (scored igraph DAG with vertex attrs year, size, group, is_birth, terminal_group, plus doc_ids), nodes (tidy tibble name, year, group, size, is_birth, terminal_group), edges (from, to, weight, documents), births (character vector of birth node names), last_year, and the de-duplicated docs_per_group.

See also

plot_trajectory_dag()

Other trajectory detection: is_flow(), sniff_trajectory_braid(), sniff_trajectory_channel(), subset.birddog_flow(), validate_flow()