Fuzzy match transform Icon Fuzzy match

Description

The Fuzzy Match transform finds strings that potentially match using duplicate-detecting algorithms that calculate the similarity of two streams of data.

For each row on the main stream, the transform compares the main stream field against values from the lookup stream and returns matches within the configured score range. Depending on the selected match mode, it can return the single best match, multiple matches as separate output rows (Top-K), or multiple matches as a concatenated list.

Supported Engines

Hop Engine

Supported

Single Threaded

Supported

Native Spark

Supported

Beam Spark

Maybe Supported

Beam Flink

Maybe Supported

Beam Dataflow

Maybe Supported

Options

General tab

Option Description

Transform name

Name of this transform as it appears in the pipeline workspace

Lookup transform

Identifies the transform that contains the fields to match

Lookup field

Identifies the field to match

Main stream field

Identifies the primary stream to match the Lookup field with

Algorithm

Identifies which string-matching algorithm to use---options include Levenshtein, Damerau-Levenshtein, Needleman Wunsch, Jaro, Jaro Winkler, Pair letters similarity, Metaphone, Double Metaphone, SoundEx, or Refined SoundEx

Case sensitive

Identifies if streams can or cannot differ based on the use of uppercase and lowercase letters---only for use with the Levenshtein algorithms

Match mode

Controls how matches are returned (see [match-mode] below)

Minimum value

Identifies the lowest possible similarity score or distance that is accepted as a match

Maximal value

Identifies the highest possible similarity score or distance that is accepted as a match

Max matches

Top-K limit for multi-match modes. Only the best matches (by distance or similarity) are kept. The default is 10. Values are hard-capped at 100 to avoid runaway memory and row explosion. Only available when Match mode is All matches (separate rows) or All matches (concatenated).

Values separator

Identifies the string that separates matches (and measure values) when Match mode is All matches (concatenated).

Match mode

Mode

Description

Closest match

Returns a single result: the best match for the current main stream value. For distance algorithms (Levenshtein, Damerau-Levenshtein, Needleman Wunsch), the best match has the smallest distance. For similarity algorithms (Jaro, Jaro Winkler, Pair letters similarity), the best match has the highest similarity.

All matches (separate rows)

Returns one output row per match, each with its own match value and measure value. Matches are ranked (best first) and limited by Max matches (Top-K). This is the recommended way to retrieve multiple matches together with their distances or similarity scores.

All matches (concatenated)

Returns a single output row where all matches are joined with the Values separator. When a Value field is configured, the corresponding measure values are joined in the same order with the same separator (for example A,B,C with measures 1,2,3). Results are also limited by Max matches (Top-K).

Existing pipelines that used the former Get closer value checkbox map as follows: checked → Closest match; unchecked → All matches (concatenated).

Algorithm Definitions

Within the Algorithm field, there are several options available to compare and match strings.

  • Levenshtein and Damerau-Levenshtein: calculate the distance between two strings by looking at how many edit transforms are needed to get from one string to another. The former only looks at inserts, deletes, and replacements. The latter adds transposition. The score indicates the minimum number of changes needed. For instance, the difference between John and Jan would be two; to turn the name John into Jan you need one transform to replace the O with an A, and another transform to delete the H.

  • Needleman Wunsch: calculates the similarity of two sequences and is mainly used in bioinformatics. The algorithm calculates a gap penalty. The aforementioned example would have a score of negative two.

  • Jaro and Jaro Winkler: calculate a similarity index between two strings. The result is a fraction between zero, indicating no similarity, and one, indicating an identical match.

  • Pair letters similarity: dissects the two strings in pairs and calculates the similarity of the two strings by dividing the number of common pairs by the sum of the pairs from both strings.

  • Metaphone, Double Metaphone, SoundEx, and Refined SoundEx: are https://en.wikipedia.org/wiki/Phonetic_algorithm#:~:text=A%20phonetic%20algorithm%20is%20an,indexing%20words%20in%20other%20languages.[phonetic algorithms], which try to match strings based on how they would sound. Each is based on the English language and would not be useful to compare other languages.

    • The Metaphone algorithm returns an encoded value based on the English pronunciation of a given word. The encoded value of the names John and Jan would return the value JN for both names.

    • The Double Metaphone algorithm has fundamental design improvements over its predecessor and uses a more complex ruleset for coding. It can return a primary and a secondary encoded value for a string. The names John and Jan each return metaphone key values of JN and AN.

    • The Soundex algorithm returns a single encoded value for a name that consists of a letter followed by three numerical digits. The letter is the first letter of the name, and the digits encode the remaining consonants.

    • The Refined SoundEx algorithm is an improvement over its predecessor. Encoded values for this algorithm are six digits long, the initial character is encoded, and multiple possible encodings can be returned for a single name. Using this algorithm, the name John returns the values 160000 and 460000, as does the name Jan.

Fields tab

The Fields tab enables you to define how to return the results of a comparison.

Option Description

Match field

Defines the name of the column that contains the comparison value (the matched lookup string). In All matches (concatenated) mode this can contain multiple values joined by the values separator.

Value field

Defines the name of the column that contains the measure for the match. The measure type depends on the algorithm and match mode:

* Distance algorithms (Levenshtein, Damerau-Levenshtein, Needleman Wunsch): integer distance for Closest match and All matches (separate rows); a separator-joined string of distances for All matches (concatenated). * Similarity algorithms (Jaro, Jaro Winkler, Pair letters similarity): numeric similarity for Closest match and All matches (separate rows); a separator-joined string of scores for All matches (concatenated). * Phonetic algorithms: the encoded phonetic value as a string.

You can also specify the list of additional fields to retrieve from the lookup stream.

  • Additional fields are available for Closest match and All matches (separate rows) (one lookup field value per output row).

  • They are also available for phonetic algorithms.

  • They are not returned in All matches (concatenated) mode for distance and similarity algorithms, because multiple matches are collapsed into a single string cell.

Output behavior

Match mode Output shape Measure (Value field)

Closest match

One output row per main stream row

Single typed measure (integer / number / string)

All matches (separate rows)

Up to Max matches output rows per main stream row (one per match), ranked best-first

One typed measure per output row

All matches (concatenated)

One output row per main stream row

Separator-joined string of measures, aligned with the match list

If no lookup value falls within the configured minimum and maximal range, the transform still forwards the main stream row with empty match and measure fields.