❄️
splink
is a Python package for probabilistic record linkage (entity resolution).
Its key features are:
-
It is extremely fast. It is capable of linking a million records on a laptop in around a minute.
-
It is highly accurate, with support for term frequency adjustments, and sophisticated fuzzy matching logic.
-
Linking jobs can be executed in Python (using the
DuckDB
package), or using big-data backends likeAWS Athena
andSpark
to link 100+ million records. -
Training data is not required because models can be trained using an unsupervised approach.
-
It produces a wide variety of interactive outputs, helping users to understand their model and diagnose linkage problems.
The core linkage algorithm is an implementation of Fellegi-Sunter's model of record linkage, with various customisations to improve accuracy.
Splink deduplicates and/or links records from datasets that lack a unique identifier.
It assumes that prior to using Splink your datasets have been standardised so they all have the same column names, and consistent formatting (e.g. lowercased, punctuation cleaned up).
For example, a few of your records may look like this:
row_id | first_name | surname | dob | city |
---|---|---|---|---|
1 | lucas | smith | 1984-01-02 | London |
2 | lucas | smyth | 1984-07-02 | Manchester |
3 | lucas | smyth | 1984-07-02 | |
4 | david | jones | Leeds | |
5 | david | jones | 1990-03-21 | Leeds |
Splink produces pairwise predictions of the links:
row_id_l | row_id_r | match_probability |
---|---|---|
1 | 2 | 0.9 |
1 | 3 | 0.85 |
2 | 3 | 0.92 |
4 | 5 | 0.7 |
And clusters the predictions to produce an estimated unique id:
cluster_id | row_id |
---|---|
a | 1 |
a | 2 |
a | 3 |
b | 4 |
b | 5 |
Splink works best when the input data has multiple columns, and the data in the columns is not highly correlated. For example, if the entity type is persons, you may have their full name, date of birth and city. If the entity type is companies, you may have their name, turnover, sector and telephone number.
Splink will work less well if all of your input columns are highly correlated - for instance, city, county and postal code. You would need to have additional, less correlated columns such as full name or date or birth, for the linkage to work effectively.
Splink is also not designed for linking a single column containing a 'bag of words'. For example, a table with a single 'company name' column, and no other details.
The homepage for the Splink documentation can be found here. Interactive demos can be found here, or by clicking the following Binder link:
The specification of the Fellegi Sunter statistical model behind splink
is similar as that used in the R fastLink package. Accompanying the fastLink package is an academic paper that describes this model. A series of interactive articles also explores the theory behind Splink.
The Office for National Statistics have written a case study about using Splink to link 2021 Census data to itself.
Splink supports python 3.7+. To obtain the latest released version of splink:
pip install splink
The following code demonstrates how to estimate the parameters of a deduplication model, use it to identify duplicate records, and then use clustering to generate an estimated unique person ID.
For more detailed tutorials, please see here.
from splink.duckdb.duckdb_linker import DuckDBLinker
from splink.duckdb.duckdb_comparison_library import (
exact_match,
levenshtein_at_thresholds,
)
import pandas as pd
df = pd.read_csv("./tests/datasets/fake_1000_from_splink_demos.csv")
settings = {
"link_type": "dedupe_only",
"blocking_rules_to_generate_predictions": [
"l.first_name = r.first_name",
"l.surname = r.surname",
],
"comparisons": [
levenshtein_at_thresholds("first_name", 2),
exact_match("surname"),
exact_match("dob"),
exact_match("city", term_frequency_adjustments=True),
exact_match("email"),
],
}
linker = DuckDBLinker(df, settings)
linker.estimate_u_using_random_sampling(target_rows=1e6)
blocking_rule_for_training = "l.first_name = r.first_name and l.surname = r.surname"
linker.estimate_parameters_using_expectation_maximisation(blocking_rule_for_training)
blocking_rule_for_training = "l.dob = r.dob"
linker.estimate_parameters_using_expectation_maximisation(blocking_rule_for_training)
pairwise_predictions = linker.predict()
clusters = linker.cluster_pairwise_predictions_at_threshold(pairwise_predictions, 0.95)
clusters.as_pandas_dataframe(limit=5)
We are very grateful to ADR UK (Administrative Data Research UK) for providing the initial funding for this work as part of the Data First project.
We are extremely grateful to professors Katie Harron, James Doidge and Peter Christen for their expert advice and guidance in the development of Splink. We are also very grateful to colleagues at the UK's Office for National Statistics for their expert advice and peer review of this work. Any errors remain our own.