AI Glossary / Retrieval and knowledge systems / Intermediate
Match rating approach
Definition
The Match rating approach (MRA) is a rule-based method for comparing names through compact phonetic encodings and a rating procedure. It can accept two names as candidate matches even when their encoded strings are not identical. The method is associated with Western Airlines and is useful to understand as an early approach to approximate name matching. Its match decision is not proof that two records identify the same person.
What it means
MRA has two distinct stages: producing a shortened representation of each name and comparing the resulting representations. This distinction matters when choosing software. Comparing two stored codes for exact equality is not necessarily equivalent to applying the full rating procedure.
The result is a rule-based candidate decision, not a calibrated probability. A positive result does not mean a particular percentage chance of identity, and a negative result does not establish that the people are different. Names are only one part of the evidence available in record linkage.
How it works
In the Apache implementation, preprocessing normalizes case and selected punctuation and accents. Encoding removes internal vowels, reduces doubled consonants, and retains the first and last three characters when the result exceeds six characters.
Comparison rejects code-length differences of at least three. Otherwise, a length-based threshold is applied to a rating derived from unmatched characters after comparisons from both ends. The implementation returns a candidate decision. Its comparison method name can be misleading: the source performs this rating procedure rather than simply testing code equality. Applications should document the library and settings used.
Three practical examples
- A booking-data review can use MRA to suggest name pairs for inspection when a spelling variation prevents exact matching.
- A historical-record search can compare candidate spellings while showing users the original names and record references.
- A data-quality experiment can compare MRA candidates with exact-match candidates to measure additional useful discoveries and additional review work.
Think of it this way
Think of two clerks shortening names on index cards and then comparing what remains from the beginning and the end. They can flag cards that deserve a closer look without declaring the cards interchangeable. The analogy stops at human interpretation: MRA follows fixed rules and has no knowledge of the people, their relationships, or the circumstances behind a spelling.
Worked example
Consider an original manual trace using the reviewed Apache rules. Carter becomes CRTR after its internal vowels are removed. Carton becomes CRTN. Neither result needs shortening or duplicate-consonant removal. The two codes each have length four, so their length difference is zero and their combined length is eight.
For that combined length, the minimum rating is three. Comparing the two codes from both ends removes the shared C, R, and T positions. One unmatched character remains in each code: R in CRTR and N in CRTN. The rating is six minus the longer remaining length, giving 6 - 1 = 5. Five meets the threshold of three, so the comparison accepts this pair as a candidate.
The encoded strings CRTR and CRTN are different. Exact code equality would reject them, showing why it is important to distinguish the encoder from the complete comparison procedure. This calculation follows source rules; it is not a reported execution of the Java library.
Now suppose a fictional review collection contains four Carter-query candidates accepted by the procedure, including a Carton record. Independent record review finds that only one of the four belongs to the intended person. Candidate precision is then 1/4, or 25 percent. Nothing about the rating of five implies 5/6 identity confidence. The fraction measures this deliberately constructed review set, while the rating is an internal rule quantity.
If a trusted contextual filter removes two unrelated candidates while retaining the relevant one, precision becomes 1/2. That improvement comes from additional evidence, not from changing the interpretation of the MRA score.
Illustrative case study
Fictional case: a travel-support team discovers that its imported customer index uses exact equality of MRA codes, while an older search tool uses the full comparison method. Staff assume that both tools must return identical candidates because they share the same algorithm name.
The team writes a compatibility test containing names with unequal codes that pass the rating rule. It labels the two search modes accurately and measures the difference in review workload. Account records remain separate until the normal identity-resolution process supplies sufficient evidence.
Common misconceptions and mistakes
Do not interpret the word rating as a probability or a universal distance measure. Do not assume that matching is limited to identical codes, or that every library with a similarly named method follows identical preprocessing.
Deleting parts of a name inevitably loses information. Short names, unfamiliar scripts, and unexpected punctuation deserve explicit tests. Missing values should remain distinguishable from real names. A method that works acceptably for one collection may introduce too many collisions or misses in another collection with different spelling conventions.
Business use
MRA can serve as a candidate generator or an additional comparison feature in a larger record-linkage workflow. Its practical value depends on the cost of reviewing extra candidates and the cost of missing genuine variants.
Build a labeled evaluation set with representative spelling differences and difficult nonmatches. Report candidate recall and precision separately, along with the number of pairs requiring review. Preserve the original names so a compressed code never becomes the only surviving evidence. Record the implementation version and test compatibility before combining indexes produced by different systems.
People also ask
Can unequal MRA codes still match?
Yes. The full comparison procedure can accept unequal encodings through its rating rule.
Is MRA a trained machine-learning model?
The method described here is a fixed rule-based procedure, though its outputs can be used within a learned linkage system.
Does an accepted comparison authorize merging records?
No. It supplies a candidate relationship that requires the application's broader evidence and decision rules.
Check your understanding
In the Carter/Carton example, does a rating of five mean an approximately 83 percent probability of shared identity?
Answer: No. Five is a rule-derived rating compared with a threshold. It is not a calibrated probability and cannot establish identity.
Related terms
- Soundex
- Metaphone
- Record linkage
- Candidate generation
- Precision
- Recall
- Phonetic encoding
- Entity resolution
- String similarity
Learn next
Compare an exact-code search with the full rating procedure on a small labeled set of invented names. Count additional candidates and determine which are useful. Explain why the denominator for candidate precision changes when extra nonmatching records are retrieved, even if the relevant record remains present.
More in Retrieval and knowledge systems
- Long tail - The long tail is the large collection of relatively infrequent items or events beyond a small group of highly popular ones in a ranked distribution.
- Low-dimensional embedding - A low-dimensional embedding represents an item using relatively few numerical coordinates while aiming to retain relationships useful for a task.
- Macroaveraging - Macroaveraging combines evaluation results by calculating a metric separately for each class or other defined unit and then taking their unweighted...
- Microaveraging - Microaveraging combines evaluation evidence across classes or labels before calculating the final metric.
- Moving average - A moving average is an average recalculated over a window that moves through a sequence of observations.
- Music alignment - Music alignment is the process of identifying corresponding musical positions across different representations or performances of a work.
Sources
Primary implementation source: Apache Commons Codec MatchRatingApproachEncoder source: https://commons.apache.org/proper/commons-codec/apidocs/src-html/org/apache/commons/codec/language/MatchRatingApproachEncoder.html . The Carter/Carton trace and review-set arithmetic are original manual examples. Historical attribution follows the implementation documentation; the original airline specification was not consulted.
Previous: Master sample Next: Matched filter