Appraise the content coherence of a trajectory flow

Scores how content-coherent a sniff_trajectory_braid() / sniff_trajectory_channel() partition is, using a silhouette over an independent content signal (shared references or keywords) the detectors never optimized. Each group-year node is profiled from its documents' features; node-to-node distance is 1 - Salton cosine; the silhouette is computed at two cluster resolutions: trajectory (each backbone/tributary chain) and final_group (each central + its tributaries).

Usage

sniff_trajectory_coherence(
  flow,
  content,
  profile = "incremental",
  signals = names(content)
)

Arguments

flow

A birddog_flow (from sniff_trajectory_braid() or sniff_trajectory_channel()).

content

Named list; each element a data frame with columns document_id and feature (long, one row per document-feature pair). The element name labels the signal in the output (e.g. coupling, keywords).

profile

"incremental" (default: profile a node from the papers new to its year, stripping cumulative carry-over) and/or "full" (all of a node's papers).

signals

Which names(content) to score (default all).

Value

An sniff_trajectory_coherence object: a list with nodes (one row per node x signal x profile x resolution) and summary (one row per signal x profile x resolution, with mean_sil, n_nodes, n_singletons, n_excluded, coverage).

See also

sniff_trajectory_comparison(), sniff_trajectory_braid(), sniff_trajectory_channel(), validate_flow()

Other flow utilities: sniff_trajectory_comparison()