Overview of Cluster Clean

Cluster clean enables users of Cloud Dataprep by TRIFACTA® INC. to standardize values in a column by clustering similar values together. Using one of the supported matching algorithms, Cloud Dataprep by TRIFACTA INC. can cluster together similar column values. You can review the clusters of values to determine if they should be mapped to the same value. If so, you can apply the mapping of these values within the application.

Example - Multiple methods of clustering

Source:

The following dataset includes some values that could be standardized:

RowIdValues
Row01Apple
Row02pear
Row03apple
Row04pair
Row05Åpple
Row06pare

When you standardize using a spelling-based algorithm, the following values are clustered:

Source ValueNew Value
Apple
apple
Åpple
Unclustered values
pear
pair
pare

After you select the cluster of values at top, you can enter apple, in the right context panel to replace that cluster of values with a single string.

In the above, the unclustered values are dissimilar in spelling, but in English, they sound the same (homonyms). When you select the Pronunciation-based algorithm, these values are clustered:

Source ValueNew Value

pear

pair
pare
Unclustered values
Appleapple
appleapple
Åppleapple

When you select the top values clustered by pronunciation, you can enter pear in the right context panel.

Results:

The six source values have been reduced to two final values through two different methods of clustering. See below for more information on the clustering algorithms.

Source ValueNew Value

pear

pear
pairpear
parepear
Appleapple
appleapple
Åppleapple

You can apply cluster-based standardization through the Standardize Page. See Standardize Page.

Clustering Algorithms

The following algorithms for clustering values are supported.

Similar strings

For comparing similar strings, the following methods can be applied:

Fingerprint

The fingerprint method compares values in the column by applying the following steps to the data before comparing and clustering:

NOTE: These steps are applied to an internal representation of the data. Your dataset and recipe are not changed by this comparison. Changes are only applied if you choose to modify the values and add the mapping.

  1. Remove accents from characters, so that only ASCII characters remain.
  2. Change all characters to lowercase.
  3. Remove whitespace.
  4. Split the string on punctuation, any remaining whitespace, and control characters. Remaining characters are assembled into groups called tokens.
  5. Sort the tokens and remove any duplicates.
  6. Join the tokens back together.
  7. Compare all tokenized values in the column for purposes of clustering.

Fingerprint Ngram

This method follows the same steps as those listed above, except that tokens are broken up based on a specific (N) number of characters. By default, Cloud Dataprep by TRIFACTA INC. uses 2-character tokens.

Tip: This method can provide higher fidelity matching, although there may be performance impacts on columns with a high number of unique values.

Pronunciation

Values are clustered based on a language-independent pronunciation.

This method uses the double metaphone algorithm for string comparison. For more information, see Compare Strings.

Var denne siden nyttig? Si fra hva du synes:

Send tilbakemelding om ...

Google Cloud Dataprep Documentation
Trenger du hjelp? Gå til brukerstøttesiden vår.