Skip to content

Choosing a matching threshold

Fuzzy matches found by uk_address_matcher in the matcher.match().matches() table have two scores associated with them: match_weight and distinguishability.1

By default, all matches are returned, no matter how low these scores. The user must therefore decide what acceptance threshold to use for these scores.

Once you have decided on a threshold, you can set it in the SplinkStage parameters:

matcher = AddressMatcher(
    canonical_addresses=df_canonical,
    addresses_to_match=df_messy,
    con=con,
    stages=[
        ExactMatchStage(),
        PeeledAddressStage(),
        UniqueTrigramStage(),
        SplinkStage(
            final_match_weight_threshold=10.0,
            final_distinguishability_threshold=1.0,
        ),
    ],
)

What do these scores mean?

Match weight

match_weight is Splink's score for the selected candidate.

Higher values mean stronger evidence that the messy address and canonical address are the same place. Lower values are weaker matches. Negative values are often a sign that the candidate was retained only because the stage is in a very permissive mode. Raising final_match_weight_threshold filters out weak matches.

Distinguishability

distinguishability is the gap in match_weight between the best candidate and the next best candidate for the same messy address. Higher values mean the winner is clearly ahead. Low values mean the top two candidates look similar, which is often where false positives come from. Raising final_distinguishability_threshold filters out ambiguous matches.

To illustrate why we need both scores, consider the following examples.

Example 1:

Messy address:
Flat 165 Block 3 Rose Square, Hammersmith And Fulham, EG1 2EG
Candidates
Flat 165 A Block 3 Rose Square, Hammersmith And Fulham, EG1 2EG
Flat 165 B Block 3 Rose Square, Hammersmith And Fulham, EG1 2EG

These two candidates will both get a high match score due to the large amount of matching information, but will have zero distinguishability.

Example 2:

Messy address:
1 High St, AB1 2CD
Candidates
1 High Street, AB1 2CD
1 High Street, WX1 2YZ

Neither of these candidate will have a high match score because they are very short, but they will have high distinguishability due to the different postcodes.

Choosing thresholds

Unfortunately, the best thresholds to use depend on various factors such as the quality and type of your input data (e.g. businesses vs residential), and whether you are matching to a small geographical area or the whole country. There's therefore no one-size-fits-all answer to what thresholds to use.

As a starting point we recommend trying the following settings:

SplinkStage(
    final_match_weight_threshold=10.0,
    final_distinguishability_threshold=1.0,
)

But to pick optimal thresholds for your use case, it's best to used labelled data, and use a precision recall curve to decide on your threshold. See the performance and benchmarking for more details on how to do this.