Fellegi-Sunter probabilistic matching
A statistical framework for record linkage that assigns weights to matched/unmatched field agreement and computes an overall match probability for each candidate pair.
Fellegi-Sunter is the foundational algorithm for probabilistic record linkage, published in 1969 and still the basis for most modern entity-resolution systems. It handles the problem of 'are these two records actually the same thing?' when fields are inconsistent, missing, or mistyped.
For each field being compared (order ID, amount, date, customer phone), Fellegi-Sunter computes two parameters: m (probability that the field agrees given the records truly match) and u (probability that the field agrees given the records truly don't match). The ratio produces a log-likelihood weight.
Summing these weights across all fields produces a match score. Scores above a threshold are labelled 'match', below another threshold 'non-match', and in between are flagged for review. This is dramatically more robust than boolean rule matching, especially for reconciliation where order IDs drift in format across systems.
ReconPe uses Fellegi-Sunter weights in its scoring stage, with the m/u parameters learned adaptively from each organisation's feedback — so the engine gets more accurate at your specific settlement patterns over time.