Skip to contents

Rank Knowledge-Graph edges by evidence strength

Source: R/causal_kg_io.R
causal_kg_rank_edges.Rd

Collapses a edaphos_causal_kg to one row per unique (cause, effect) edge and ranks the result by a user-selected metric:

Usage

causal_kg_rank_edges(
  kg,
  by = c("n_sources", "mean_confidence", "agrovoc_support"),
  alignment = NULL,
  top_n = NULL
)

Arguments

kg

An edaphos_causal_kg.

by

Character vector of ranking metrics, any subset of c("n_sources", "mean_confidence", "agrovoc_support"), in priority order.

alignment

Optional data frame as returned by causal_kg_alignment() (must have columns original, canonical, plus uri when derived from vocab = "agrovoc"). Used to compute agrovoc_support.

top_n

Optional integer — return at most this many rows. NULL returns the full ranking.

Value

A data frame with columns cause, effect, n_sources, mean_confidence, max_confidence, sources (collapsed " | "-separated), evidence (collapsed), and — when available — agrovoc_support, agrovoc_cause, agrovoc_effect.

Details

"n_sources"

Number of distinct sources supporting the edge (counted by splitting the |-separated source field on the underlying igraph edge-attribute). The single most informative signal for an LLM-extracted KG built over thousands of papers: an edge asserted by 50 papers is far more trustworthy than one asserted by 1.

"mean_confidence"

Mean LLM confidence across extractions that produced the same (cause, effect) pair.

"agrovoc_support"

Fraction of the pair's endpoints (cause, effect) that resolve to a FAO AGROVOC concept under an alignment mapping (0, 0.5 or 1). Requires either alignment to be supplied or kg already aligned; when neither is available the column is NA.

The return is a tidy data frame sorted in descending order by all ranking columns requested in by (i.e. the first by element is the primary sort key, the second breaks ties, etc.). This is usually more informative than a single-metric ranking — an edge that has both many sources AND high confidence is more trustworthy than either alone.

See also

causal_kg_alignment(), summary.edaphos_causal_kg().

Examples

# \donttest{
  kg <- causal_kg_new()
  kg <- causal_kg_add_edge(kg, "precipitation", "soc",
                            source = "Jenny 1941", confidence = 0.9)
  kg <- causal_kg_add_edge(kg, "precipitation", "soc",
                            source = "Minasny 2017", confidence = 0.85)
  causal_kg_rank_edges(kg, by = c("n_sources", "mean_confidence"))
#>           cause effect n_sources mean_confidence max_confidence
#> 1 precipitation    soc         2             0.9            0.9
#>                     sources evidence
#> 1 Jenny 1941 | Minasny 2017         
# }