What is Fuzziness?

This article applies if each source collection in your screening configurations has a fuzziness setting. Matching options for screening configurations describes the other settings that change how names are matched.

If your collections set a match threshold instead, they have no fuzziness setting. The match score decides which profiles are returned, as Understanding the Match Score explains, and Recommended Match Threshold Starting Points suggests where to begin.

Fuzziness is a matching technique that enables us to return matches when there are small variations between the spelling of your customer's name and the names referenced in risk lists and media. Fuzziness is sometimes referred to as edit distance or Levenshtein distance. Our matching algorithm uses fuzziness alongside many other techniques to match names.

Fuzziness will match entities with names that have an inserted, omitted or replaced character compared to the name you screened. Fuzziness allows up to one character difference for each word in the search term.

Fuzziness is useful in two main scenarios:

1. The customer name has been manually entered and may contain typos or spelling mistakes

2. The customer name has been transliterated into Latin script from a non-Latin script, or is written in a different script compared to a risk list or media article.

By default, our matching algorithm uses fuzziness only for the second of these scenarios: matching transliteration variants of names. This means that, by default, fuzzy name matches are only returned when the searched and matched name words are broadly phonetically equivalent (using an industry standard phonetic algorithm). 

Configuration options are available to change this default behavior such that the first scenario (typos and spelling mistakes) may also generate matches.

Understanding fuzziness percentages

Fuzziness always allows up to one character difference for each word in the search term. However, you can use the fuzziness percentage to configure how long a word must be for a difference to be considered significant.

This matters because longer names with one character different are more likely to be the same entity than shorter names. For example, the names Leederheimer and Lexderheimer are far more likely to be misspellings of each other than Lee and Lex. 

Your choice of fuzziness percentage depends on your risk-based approach and how sure you are that the names you input for searching are correct. For example, if you take the information directly from the customers' identity documents, you will have greater confidence in it than if customers input it themselves, which is more prone to error.

Names in non-Latin scripts

Before matching, a name written in a non-Latin script, such as Cyrillic or Arabic script, is transliterated into Latin script. By default, fuzziness then matches the spelling variants of that transliteration that sound the same.

For example, a search for Дмитрий Смирнов is transliterated as Dmitry Smirnov, and can match profiles spelled Dmitri Smirnov, Dmitriy Smirnov or Dmitrii Smirnov.

Names in Chinese characters are also matched in their own script:

• A search term in Chinese characters matches the same characters on a profile exactly, whatever your fuzziness setting. Traditional and simplified characters are treated as the same.

• A search term in Latin script is not converted into Chinese characters. It can match a Chinese name on a profile through that name's Latin form.

• A name recognized as a Chinese person name is matched without spelling differences, so Zeng does not match Zheng.

• The given name can be written as one word or two, so Zhang Weiming matches Zhang Wei Ming.

A person name is recognized as an Arabic name when it contains an Abd al block, such as Abd al Rahman, or a common Arabic given name with al, el, bin, ben or ibn. For these names, a word missing from the profile name can be the first or the last word of your search term.

PEP data is collected in the language of its official source, so a profile can show a name in Arabic, Hebrew or Thai script, for example. To have the Latin form of a name added to a profile, report the profile to Technical Support.

What's the difference between 0% fuzziness and exact match?

There are several differences between 0% fuzziness and an exact match. Setting fuzziness to 0% only affects the edit distance matching behavior described above.

Exact match also affects the following, in addition to setting fuzziness to 0%:

• Disables all pre-processing. For example, honorifics or suffixes like Mr, Ms, Dr or PhD are matched exactly.

• Disables all other forms of inexact name word matching such as equivalent names and phonetic matching except for word order variations and also-known-as (AKA) matching.

• Does not allow for extra words to be added. 'John Smith' won't match 'John Williams Smith'.

• Disables fuzziness for year of birth. When fuzziness is between 10% and 100%, we allow a one-year difference in year of birth.

All of the above matching behaviors are enabled by default and are disabled by using the exact match setting.

What is the impact on false positives? 

Name matching is inherently probabilistic and the options described above enable you to trade off greater aversion to the risk of missing inexact name matches against the operational impacts of a higher number of false positives. 

To optimize this trade off, we have capped general edit distance matching at one character per name word. This allows for the overwhelming majority of spelling errors and typos while controlling the number of false positives. This does not mean that only single edit distance variations are considered matches. As noted above, we additionally use many other methods to match equivalent names, phonetically equivalent words, abbreviations, hypocorisms and more. 

The ComplyAdvantage matching algorithm has been tested extensively (both internally and by independent consultants) across different names and name variations in our database.