Score candidate pairs — np

Collapses the per-field similarities of an np_compare() result into a single score in [0, 1]. Three methods are provided so the "best" combiner can be chosen empirically against a labelled training set:

Usage

np_score(
  pairs,
  config = np_config(),
  method = c("weighted", "hier", "em", "model"),
  model = NULL
)

Arguments

pairs: An np_pairs frame from np_compare().
config: An np_config().
method: One of "weighted", "em", "model".
model: A fitted model with a predict() method, required for method = "model". Must accept a data frame of the similarity columns.

Value

pairs with a numeric score column added (and method raw output).

Details

"weighted" — normalized weighted sum using the profile weights in config. Transparent, no training required. The default.
"hier" — name similarity plus a hierarchical geo sub-score that takes the strongest confirmed location granularity (ZIP+4 > street > ZIP5 > PO box > ZIP3 > city > state) via a max, not a sum. Correlated address fields are not double-counted, and a wrong city cannot pull down a ZIP-confirmed match. Degrades gracefully as geo fields go missing. See np_default_geo_weights().
"em" — unsupervised Fellegi-Sunter weights via reclin2::problink_em(); score is the match posterior. Learns weights from the data.
"model" — apply a fitted classifier (glm, ranger, xgboost, ...) supplied in model to the per-field similarity columns; score is its predicted match probability. This is the hook for supervised optimization.

The chosen method's raw output is retained (weight_em, p_model) so methods can be compared side by side before one is adopted.