Generate and compare candidate pairs — np

Blocks the query against the reference, then scores each candidate pair on the fields implied by the detected (or supplied) geographic profile. reclin2 does the blocked pairing and per-field string comparison; the result is returned as a plain data frame so every downstream npmatch stage works on ordinary columns.

Usage

np_compare(
  query,
  reference,
  config = np_config(),
  block = "state",
  profile = NULL,
  candidates = NULL,
  name_freq = NULL,
  token_idf = NULL,
  addr_freq = NULL
)

Arguments

query, reference: Normalized np_query / np_reference frames.
config: An np_config().
block: Character vector of canonical fields to block on when candidates is not supplied. Defaults to "state".
profile: Optional profile name to force; otherwise np_detect_geo().
candidates: Optional np_blocks from np_block(). When given, these candidate pairs are compared instead of blocking on block — the scalable path (name-token / progressive blocking).
name_freq, token_idf: Optional pre-computed reference tables from np_name_freq() / np_token_idf().
addr_freq: Optional np_addr_freq() table for the reference. Used to read a suite match in context: addr_shared, addr_units, addr_unit_shared and addr_note are added to the result, and np_score(method = "hier") credits a matching suite only where the address is genuinely multi-tenant. Computed from reference when NULL, which is wrong if reference is a slice of the real corpus — build it once from the whole reference and pass it in (as np_cascade() does).

Value

A data frame of candidate pairs, one row per (query, reference) pair, with class np_pairs. Columns include .x, .y, .id, .ein, a similarity column per compared field, plus carried veto features. Two columns report the suite, both three-valued (1 agree, 0 differ, NA either side names no unit): geo_unit compares the unit identifier with the designator stripped, which is the column np_score() credits and the reason STE 320 and # 320 agree; geo_unit_same_kind says whether the two sides name the same kind of place, so a reviewer reading a floor against a suite is not told they are the same office.

Details

The name is always compared (on name_key). Geographic comparison fields are taken from the profile. Name-feature columns needed by np_veto() are carried through for both sides with _x / _y suffixes.